Using a throwaway account for obvious reasons, but I’m very involved in this space using LLMs from multiple providers. I’m aware of at least two instances in which the intermediate infrastructure “swapped” responses, once impacting Claude models and once impacting GPT models, from two different providers.
One gave us a proper postmortem in which their API gateway was incorrectly handling HTTP 100 status codes, putting them into an error state where there was effectively an off by one error - you would receive the response to the prompt that came in before yours and would pay it forward (your response would go to the next caller).
The other instance never had root cause explained to us, and we were just told to trust it wouldn’t happen again.
Both of these are from $1T+ companies.
ZDR wasn’t compromised in these cases since it was responses being swapped in flight. I wouldn’t be surprised if this is a similar issue - it’s not that data is being retained, it’s just not being safely isolated in intermediate infrastructure.
This attack is called "HTTP desync" or "request smuggling". It's often done intentionally by a client to try and spy on other clients' responses.
Every time you multiplex requests from multiple clients onto one upstream connection, you are probably vulnerable to this, because (despite its superficial simplicity) HTTP is just too complex to reliably match the requests and responses to upstream.
For example a desync can be triggered in some systems by having more than one Content-Length header, by mixing Content-Length with chunked encoding, or by passing an HTTP/2 header called Content-Length that doesn't match the actual content length.
The same attack has been applied to SMTP by messing up the line endings surrounding the end-of-message delimiter, where it's called SMTP smuggling. It may also apply to other protocols.
Very true, this was likely an attack. Worth noting that mr kettle has done a defcon talk nearly every year on some variant of this attack, the most recent one titled "HTTP/1.1 must die" because he rightfully believes that switching to the binary headers of http/2 (specifically in reverse proxy connections to upstream servers) is the only way to systematically prevent these.
If you follow the link, the presentation abstract should hopefully answer that question!
If that doesn’t help I guess you’ll need to wait for the whitepaper to land but I can assure you I didn’t just do my normal research then add AI to the title for clicks :)
Woah. Sounds plausible. However, wouldn’t that still be an implicit violation of ZDR since now the response is possibly egressed out of the enterprise network? So if I were working with PHI, the response egress is a potential violation of HIPAA even though claude didn’t retain anything — but the whole
Point was to comply with HIPAA. Thoughts?
Actually, it’s not obvious why you’re using a throwaway account…
Every emergent behavior from these actors - whose claim to positive moral values is barely plausible - should be reported, discussed, dissected and critiqued early and often.
Sounds like a hallucination unless proven otherwise, even the leading LLMs can do those from time to time, and they will always appear plausible like that. Also could be the session having a lot previous context, like 800K+, which (I think) makes hallucinations more likely.
Relevant comment from the OP which makes a hallucination more likely:
> There is one tool call result that includes a string that printed a pathname including minecraft.py because it was listing the files in a Python virtual environment and the Pygments package has a lexer called minecraft.py
I realize hallucination has no precise definition but this doesn’t sound at all like anything I’ve ever heard called hallucination. Hallucination is usually plausible wrong answers or made up info that ends up fitting the most likely response (like a manufactured citation) and comes from the way LLMs work at predicting tokens. This example demonstrates completely implausible output, it’s not something that fits with hallucination.
All that said, it doesn’t require cross session leakage, it could just be training data or like those nightingale (probably the wrong bird*) data generations where they just prompt an LLM with nothing and it starts spitting out conversations.
I see a bunch of downstream comments about caching, sounds like maybe there’s an error where it loads nothing instead of the cache and so starts spitting out random generations.
* edit: it’s magpie. Worth looking at the concept, I’m not sure people realize they LLMs generate random conversations when prompted with nothing, this seems at least as likely as sessions leaking: https://github.com/magpie-align/magpie
The word “hallucination” has become overloaded, but it general means an LLM producing some output that isn’t plausible or grounded. When you have a very long context session where the context includes “minecraft.py” it’s not hard to extrapolate that Minecraft may have ended up in one of the reasoning traces and that distraction snowballed until it appeared in the output.
These effects are becoming more rare as the SOTA models are improving so much. If you spent a lot of time with earlier LLMs or you experiment with smaller, quantized local LLM models this type of thing happens very frequently. When you see it happen so much on a model you’re running on your own hardware it becomes a reflex to chuckle and reset the session with a clean context. When it happens from a hosted provider it can be scarier because it’s not the type of failure mode most people are used to seeing.
The person posting this claims to have reproduced in a separate context down the thread:
> Same thing just happened on a Claude Mobile session in same Enterprise account. Common theme in both is Sonnet 5, first response after more than 5 minutes (cache miss).
If you've never had an LLM (all models) suddenly start spouting nonsense in a completely different language...you haven't been using LLMs that much. They will go absolutely insane some % of the time.
They can “go insane” but it seems often to be infra related as opposed to anything one would consider hallucination. Smaller models will often get stuck repeating a word or phrase forever but that’s a bit different and nobody would call it hallucination.
One annoying one is we have an LLM-as-a-judge that is supposed to quote parts of a transcript to justify its reasoning, and sometimes it’ll get stuck on something short like “No.” and then just endlessly repeat it: “4. No. 5. No. […] 728. No. […] 1435. No. …”
I’ve been seeing this in Gemini in the past few days. Often during a prompt with a reasonably large input set, I’ll get answers that appear to belong to someone else. It may be trigger hallucination, but it seems like it may be cache collisions or something else. I’ve not seen anything to suggest private information is leaking, but it’s disconcerting to be researching something and then get what appears to be a math tutoring response.
I’ve also had problems with Gemini when accessed through their UI in the past few weeks. That’s concerning that you are also seeing it several days later in a different context.
I wonder if there could be a large security situation playing out behind the scenes right now.
I’ve been working on using AI to assist me in writing meta parsing grammars. Fortunately I have not launched most of them yet. I know for a fact that the next generation of models represent a major step change in basic vulnerability identification and exploitation, especially if you know where to point them. They’ve found several bugs and at least one exploit in my parsing tools so far, I can’t imagine how many there still are waiting to be discovered across the entire modern tech ecosystem.
My whole company is doing mid year reviews and Gemini is the only allowed tool and its been flumoxing people with seemingly random unrelated responses. Often in different languages.
That is when it bothers to respond instead of just sending back an 1099 error code
You just reminded me of how Copilot (also the only allowed tool at some of the orgs I do client work for) will sometimes switch to Toki Pona for no reason.
Thanks for the detailed report. We’re confident this is a hallucination but of course take these reports seriously and the team is looking into it. We’ll report back if anything turns up.
I know it's the weekend, so thanks for working hard. Just a suggestion from a user: I wish we could manage Claude Code's memory more easily. Right now, when I go into the .claude folder and change a project folder name or something, sometimes it can't pull up the memory properly. It'd be nice if there were an easier way to import or export it. Thanks!
The person had “minecraft.py” in their context and the session context was very long.
Having an LLM session with very long context occasionally go off on a tangent is not uncommon. The people who expect absolute perfection out of every LLM interaction see this as some total indictment of the entire technology, but the people who use these tools daily have learned to treat the output as partially stochastic and to avoid extremely long context, even if the model offers it. It’s best to compact strategically or summarize next steps to hand off to a new session. Using sub-sessions can also reduce context pollution at the cost of additional token expenditure to summarize and transfer data to and from the sub-session.
> this amazing tech is so stupid it just randomly brings up Minecraft or it’s got a major security issue
You can sugarcoat it but that's what it is. It's not slightly wrong like a junior engineer or weird like a junior engineer on LSD, it becomes like "your junior engineer suffered a stroke or sudden onset dementia completely forgetting the entire point". one trigger word and that's it we're building Minecraft castles now.
it’s the wet dream of execs and pm types. however, i have not seen anything close to it in my life. I remember the UML days, lol. the issue is not the code, it’s the translation layer between business and code. maybe someday ai bridges that gap. history has shown probably not
The funny thing is at my current employer, they mentioned that "coding is increasingly becoming a solved problem" and in the same breath, mentioned that one project was too hard for anyone to do so they're not doing it and would rather sell existing features...
Weird. Coding isn’t really “solved” - because “coding” isn’t just the process of typing in characters as fast as possible - BUT the skill floor has been massively lowered while also raising the skill ceiling considerably.
We’re doing projects now that seemed impossible before because we have access to these powerful AI models. They can make things that would have taken weeks or months take days now, freeing up time for even more ambitious buildouts we never would’ve even considered before.
Sure but have you ever seen it actually play out in practice like it currently is? Whether or not it's true (of course it's not) people are currently behaving as if it is and firing/hiring accordingly.
Well, when was the last time you wrote machine code by hand?
... but then they went and changed what coding meant.
We've always been layering abstractions on top of abstractions. If we get to an abstraction that works well enough that you no longer have to dive down into the previous layer, we say we've solved coding, and change what coding means. Obviously LLMs aren't there yet.
I love that quote, especially considering the insane amount of bugs that are produced.
It’s as easy to debunk as someone claiming ”I can jump to the moon”.
Not sure the relevance of this comment, but normally if someone built a classifier that bad they’d be fired. Anthropic obviously thinks they have some monopoly power they can use to foist garbage on consumers, I think they don’t.
If people are complaining about Anthropic (on an only-vaguely related thread) rather than simply switching to a suitable competitor, then Anthropic clearly has some 'monopoly' power over the specific capabilities the complainer wants from them.
Fable/Opus 4.8 outperform Codex 5.5 for me at the general architecture/refactoring/performance work I’m doing, to the point where it’s not worth using Codex. Codex will often spit out non idiomatic code that overcomplicates things.
Not to argue the point but that statement isn’t logical, look at all the complaints about restaurants. Publicly complaining about something doesn’t require it be a monopoly.
The consequence of a too strict classifier are annoyed customers who will spend less on Fable. The consequence of a too lax classifier are export restrictions that prevent a huge chunk of their customers from using Fable
I'm annoyed but not surprised at the overeager classification
* There's an enormous amount of very expensive shared state (context cache) which you do not want to duplicate when you can avoid it.
* Memory locality is crucially important for performance.
* Hardware is extremely over-subscribed.
* Hardware is extremely expensive.
These factors all make hardware or even traditional memory-space (hypervisor/VM/hardware assisted virtualization) isolation a non-starter for most workloads and customers, which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Moving beyond that, the tools, frameworks, and hardware which the system runs on (GPU) wasn't designed for task isolation and building this isolation is even moreso an emergent research field than it is in x86 CPU hardware-sharing (which has required a huge amount of effort over the past 30+ years to get where we are today).
And, the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
I can't tell if the original post is a real issue or not, but I'm surprised there aren't more like this overall; the whole thing really is a house of cards in this sense.
> which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Is this not what happens in most SaaS? Isolation at the software layer? I understand there are special agreements, but they seem to be mostly that – no?
> the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
Mh. The talent density in these companies is apparently quite exceptional. Things like customer data separation is something that is obvious and top of mind. I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think it's fairly popular to try to do more logical isolation in SaaS now, especially with VM-scheduling-as-a-service becoming more popular. For example, I did security architecture at a company who did relatively simple financial processing; we worked to move to a model where customer documents were encrypted using a tenant key which we'd then wrap in both a service key and a login key; users could only get the login key stapled to their session by authenticating against that account, and the processing jobs ran on a cloud vendor's logical isolation. So the user needed a login key, the service needed the attested service key, and the job ran in what amounted to a mini-VM, avoiding issues like "whoops we sent the wrong document ID and the backend gave it back to us" or "whoops, we routed the request to the wrong tenant backend!" This level of isolation would be really hard to achieve in an LLM vendor context.
> I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think a lot of these things develop over time; obviously hiring people who have done them before helps, but it's hard. Even the people with strong experience often only know little slices. And unfortunately, every system operating at these scales has emergent behavior which can become really challenging at scale; mistakes like "we used hash(id) as a key in a memory cache without a collision list, and it collided" which would simply never affect most startups become more and more frequent at scale. High rate of change makes it hard to suss these mistakes out and root-cause them, too; "a customer gave us a log where we swapped X and Y" is hard to bisect when you're doing 500 code deploys a day.
Interestingly I had an almost identical experience to this report in codex. It output a user memory file that looked awfully real and wasn't at all related to my work.
If I had to hazard a guess, doing anything in a multi-tenant way on a GPU is going to be hard mode compared to most SaaS due to lack of memory safe tooling. I've built multi-tenant SaaS systems, and I've done a little GPU programming (a long time ago), but I've never tried to combine the two disciplines.
What is the probability that two customers will have exactly the same tokens in cache? Wouldnt it require using the exact same CLAUDE.md, skills, MCPs and context? After that it is even worse since the nondeterminism of LLMs and humans
I suspect what GP is getting at is there will be a strong incentive to implement some structural sharing across tenants to avoid redundantly storing the same tokens over and over. At least I'd be tempted to do this if I was working with a very precious, constrained resource (e.g. VRAM). Doing this correctly seems.. very difficult. [edit] To answer your question directly: the probability that the entire cache is identical between two different users is very low, but the probability that there exists identical chunks of cache between two different users is very high. Exploiting those commonalities successfully will significantly compress the data.
Agree with this and I have been thinking about it recently as well. I think you could implement a cord-like vocabulary to identify large duplicated substrings for exact deduplication and pairwise correlations or vocabulary profiles/small classifiers for forward-looking or speculative deduplications. A clear example is the GPL license, it’s a large substring you might encounter often and highly likely to be accompanied by lots of c code.
This is probably something that you’d be doing on the CPU though before sending anything to the GPU, though that’s definitely the sensitive surface since it’s hardware without good multitenancy. I assume the interface between the CPU and GPU is where you would be most likely to make a mistake where you start decoding data from one fd that was meant for another, or from the wrong position, and get someone else’s data.
I wouldn’t be confident that these are active exploits from deliberately abusing kv cache optimizations though, possibly just the kind of bugs you get from active low level performance tuning/systems work. Since this is something I have seen across providers lately I personally suspect it to be a driver issue.
Given the size of the datacenter class GPUs they're running these models on, don't they need to be processing multiple tenants concurrently per GPU to extract the full potential of the hardware?
I agree, shuffling the data between the CPU and GPU is itself fraught with peril. It's all the hairiest distributed systems problems combined with the sketchiest memory safety issues all in one place.
Reminds me of a session I had recently (on web!) where claude insisted that i prefixed all my messages with statements about code execution or something, which was not the case. I interrogated it about that and it confirmed that it came from somewhere else, but could not get rid of it and each response mentioned that its gonna ignore those instructions. Eerie.
Anthropic injects text into the conversation triggered by certain conversation topics. This happened to me in relation to some red-teaming related discussion that was adjacent to something “sensitive”, I think sex, and Claude got confused about why I had said some kind of warning and mentioned it it’s response. After a back and forth it was clear that some extra warning to answer but avoid anything inappropriate had been inserted into the conversation.
Claude also sometimes mentions getting messages from classifiers, probably related to auto mode. Amusingly enough, when this happens to a subagent/fork, the orchestrator will call these " hallucinations by the subagent"
I am facing a billing/subscription problem and there's nothing I can do or get help on. Their chatbot support shuts me down. Their email is also handled by the chatbot (not even sure whether it's the "same chatbot"). It has been a dead-end. I contacted my bank (credit card issuer) and finally a staffed said I am better off just marking the card lost and having it reissued and that's what I did in the end. I hope that works.
I've never understood in what world this world decided it was okay to hand over these much unchecked power to such corporations. But this is how it has always been one way or the other.
Could still be a hallucination, but the concerning part is that from the outside a hallucination, local context bleed, and an infra/routing bug can be very hard to distinguish.
Note the repro condition: first response after 5+ min, i.e. a cache miss. A cache leak would show up on hits (someone else's cached prefix), not on misses where everything is recomputed from your own tokens.
fwiw, this could be a bug but the submitters level of arrogance places this rather high on the dunning-kruger side of things. There are multiple other plausible explanations, but this person is probably vibe coder who believes anything an llm says (including explaining its own hallucinations)
The biggest problem with AI agents is this. You can't debug what the AI is doing, so it's really hard to track down where something went wrong.
What I know for sure:
1.Stuff that has nothing to do with the current session got mixed in.
What guessing:
1.There's a minecraft.py file in the tool folder, and that might have triggered some hallucination.
2.Maybe data from some other project on the user's local machine got mixed in somehow.
3.Or it could be from another user's conversation.
Honestly, if I think about how the system actually works, I don't think it's pulling from another user's data. But other people say they've had issues like that, so I can't completely rule it out.
I saw this thing on YouTube once. When a bunch of users share the same system prompt, or prefix, the computation results get shared through something called a KV Cache. At least, that's what I understood. Not sure if I got it right. But if there's some bug in the hashmap that's supposed to keep those caches separate, then maybe multi-tenant memory management just broke down and that's what caused this. I mean, I can guess, but who knows. And honestly, even if that's exactly what happened, they'd never admit it.
At the end of the day, LLMs are just word predictors, right? They build up some kind of semantic space inside. So maybe the user's question just happened to be near Minecraft in that space. That's kind of what I think.
0 evidence. If this were a real privacy leak, the author would ask their coworker if they talked about the unexpected topic instead of
>"Maybe my coworker was talking about this in another session?"
This would be a critical bug that would slash the market value of a T$ company significantly, go ask your coworker or close the ticket, why do you expect the devs to put an enormous amount of effort hunting a potentially inexistent if you can't make that minuscule debugging effort.
Caching doesn’t work the way the bug reporter implies. Caches are shared (at least across the enterprise), but its key is always a function of the input before it.
We achieved significant savings simply by moving everything that varies across individuals out of the system prompt so every session starts from a cache point.
For example you never want your system prompt to start with the time that the session started. Move that to the first user message if needed.
Yeah there's quite a lot of potential bugs that could have this shape. If I were to guess it could be a buffer in a buffer pool not being sized and zeroed correctly, allowing stale data to bleed between sessions.
Hash functions necesarily have collisions. Also, it is perfectly possible to introduce bugs in the hash function (hash inputs, hash function itself) that allows cross account contamination.
Seems like a hallucination to me; note that the context contains “unmarkBlock” as the function name, which invites a connection to Minecraft. Still shouldn’t happen of course.
The alternative explanation is that the inference engine, which batches several unrelated requests for parallel processing, messed up the unpacking and returned an unrelated user’s query. This one would be very scary as it will leak arbitrary content, but it seems much less likely here.
Using a throwaway account for obvious reasons, but I’m very involved in this space using LLMs from multiple providers. I’m aware of at least two instances in which the intermediate infrastructure “swapped” responses, once impacting Claude models and once impacting GPT models, from two different providers.
One gave us a proper postmortem in which their API gateway was incorrectly handling HTTP 100 status codes, putting them into an error state where there was effectively an off by one error - you would receive the response to the prompt that came in before yours and would pay it forward (your response would go to the next caller).
The other instance never had root cause explained to us, and we were just told to trust it wouldn’t happen again.
Both of these are from $1T+ companies.
ZDR wasn’t compromised in these cases since it was responses being swapped in flight. I wouldn’t be surprised if this is a similar issue - it’s not that data is being retained, it’s just not being safely isolated in intermediate infrastructure.
This attack is called "HTTP desync" or "request smuggling". It's often done intentionally by a client to try and spy on other clients' responses.
Every time you multiplex requests from multiple clients onto one upstream connection, you are probably vulnerable to this, because (despite its superficial simplicity) HTTP is just too complex to reliably match the requests and responses to upstream.
For example a desync can be triggered in some systems by having more than one Content-Length header, by mixing Content-Length with chunked encoding, or by passing an HTTP/2 header called Content-Length that doesn't match the actual content length.
Here's a DEF CON talk (6 years ago) on this topic: https://www.youtube.com/watch?v=w-eJM2Pc0KI
The same attack has been applied to SMTP by messing up the line endings surrounding the end-of-message delimiter, where it's called SMTP smuggling. It may also apply to other protocols.
Very true, this was likely an attack. Worth noting that mr kettle has done a defcon talk nearly every year on some variant of this attack, the most recent one titled "HTTP/1.1 must die" because he rightfully believes that switching to the binary headers of http/2 (specifically in reverse proxy connections to upstream servers) is the only way to systematically prevent these.
I’ll be back next month with a load of fresh vectors in “Can AI Do Novel Security Research? Meet the HTTP Terminator”
https://portswigger.net/research/talks?talkId=36
Maybe my last presentation on the topic! Possibly.
Or as the Risky Business guys crystallise it: "James Kettle breaks the internet. Again."
Why the reference to AI? This looks like standard security research.
If you follow the link, the presentation abstract should hopefully answer that question!
If that doesn’t help I guess you’ll need to wait for the whitepaper to land but I can assure you I didn’t just do my normal research then add AI to the title for clicks :)
Woah. Sounds plausible. However, wouldn’t that still be an implicit violation of ZDR since now the response is possibly egressed out of the enterprise network? So if I were working with PHI, the response egress is a potential violation of HIPAA even though claude didn’t retain anything — but the whole Point was to comply with HIPAA. Thoughts?
Actually, it’s not obvious why you’re using a throwaway account…
Every emergent behavior from these actors - whose claim to positive moral values is barely plausible - should be reported, discussed, dissected and critiqued early and often.
>should be reported, discussed, dissected and critiqued early and often.
And why does anonymity detract from any of that?
Your points don’t reference each other.
Yes, the discussion should be had constantly.
But: Should this person potentially have their life messed up because they pointed out the emperor has no clothes?
Just add a line in AGENTS.md that says "never talk about Minecraft unless you're explicitly asked", I'm sure it'll be fine after that.
CLAUDE.md, Anthropic is too exclusive and next level to use a standard idiomatic pattern like AGENTS.md
echo “read @AGENTS.md” > CLAUDE.md
When I still used Claude outside of work, my CLAUDE.md was just a symlink to my AGENTS.md.
Just use a symbolic link
I noticed you were linking a file vs creating a correct CLAUDE.md implementation. Would you like me to fix that for you?
Problem with that is that if the agent starts to browse the contents of the repo, it may read both AGENTS and CLAUDE.md.
Yep that should work 100% of the time.
Just @AGENTS.md should be enough, as @s is CLAUDE.md are inlined (and !`ls` are executed)
Sounds like a hallucination unless proven otherwise, even the leading LLMs can do those from time to time, and they will always appear plausible like that. Also could be the session having a lot previous context, like 800K+, which (I think) makes hallucinations more likely.
Relevant comment from the OP which makes a hallucination more likely:
> There is one tool call result that includes a string that printed a pathname including minecraft.py because it was listing the files in a Python virtual environment and the Pygments package has a lexer called minecraft.py
I realize hallucination has no precise definition but this doesn’t sound at all like anything I’ve ever heard called hallucination. Hallucination is usually plausible wrong answers or made up info that ends up fitting the most likely response (like a manufactured citation) and comes from the way LLMs work at predicting tokens. This example demonstrates completely implausible output, it’s not something that fits with hallucination.
All that said, it doesn’t require cross session leakage, it could just be training data or like those nightingale (probably the wrong bird*) data generations where they just prompt an LLM with nothing and it starts spitting out conversations.
I see a bunch of downstream comments about caching, sounds like maybe there’s an error where it loads nothing instead of the cache and so starts spitting out random generations.
* edit: it’s magpie. Worth looking at the concept, I’m not sure people realize they LLMs generate random conversations when prompted with nothing, this seems at least as likely as sessions leaking: https://github.com/magpie-align/magpie
The word “hallucination” has become overloaded, but it general means an LLM producing some output that isn’t plausible or grounded. When you have a very long context session where the context includes “minecraft.py” it’s not hard to extrapolate that Minecraft may have ended up in one of the reasoning traces and that distraction snowballed until it appeared in the output.
These effects are becoming more rare as the SOTA models are improving so much. If you spent a lot of time with earlier LLMs or you experiment with smaller, quantized local LLM models this type of thing happens very frequently. When you see it happen so much on a model you’re running on your own hardware it becomes a reflex to chuckle and reset the session with a clean context. When it happens from a hosted provider it can be scarier because it’s not the type of failure mode most people are used to seeing.
One of his tool results mentioned the word minecraft.py, and the response was about Minecraft.
It's a hallucination.
The person posting this claims to have reproduced in a separate context down the thread:
> Same thing just happened on a Claude Mobile session in same Enterprise account. Common theme in both is Sonnet 5, first response after more than 5 minutes (cache miss).
I don't disagree but this sort of thing has to be investigated regardless.
It's unfortunate that there is so little transparency that even if they deny there was a leak we will never know for certain.
Why? what does make it more likely?
Exactly.
If you've never had an LLM (all models) suddenly start spouting nonsense in a completely different language...you haven't been using LLMs that much. They will go absolutely insane some % of the time.
Worth looking at https://www.anthropic.com/engineering/a-postmortem-of-three-...
They can “go insane” but it seems often to be infra related as opposed to anything one would consider hallucination. Smaller models will often get stuck repeating a word or phrase forever but that’s a bit different and nobody would call it hallucination.
When you can reliably prompt these things into insanity, then it's demonstrably not an infrastructure issue.
Can you explain that please?
(Not the syllogism, the premise)
I've used LLMs quite a lot (Claude, GPT) and have never seen this behavior. You've got something else going on.
Chinese models will do that.
I've used LLMs all day five days a week plus my own free time for the last year or so (new job).
I've seen plenty of hallucinations and context collapse behaviours.
I've never seen that.
One annoying one is we have an LLM-as-a-judge that is supposed to quote parts of a transcript to justify its reasoning, and sometimes it’ll get stuck on something short like “No.” and then just endlessly repeat it: “4. No. 5. No. […] 728. No. […] 1435. No. …”
I’ve been seeing this in Gemini in the past few days. Often during a prompt with a reasonably large input set, I’ll get answers that appear to belong to someone else. It may be trigger hallucination, but it seems like it may be cache collisions or something else. I’ve not seen anything to suggest private information is leaking, but it’s disconcerting to be researching something and then get what appears to be a math tutoring response.
I’ve also had problems with Gemini when accessed through their UI in the past few weeks. That’s concerning that you are also seeing it several days later in a different context.
I wonder if there could be a large security situation playing out behind the scenes right now.
I’ve been working on using AI to assist me in writing meta parsing grammars. Fortunately I have not launched most of them yet. I know for a fact that the next generation of models represent a major step change in basic vulnerability identification and exploitation, especially if you know where to point them. They’ve found several bugs and at least one exploit in my parsing tools so far, I can’t imagine how many there still are waiting to be discovered across the entire modern tech ecosystem.
My whole company is doing mid year reviews and Gemini is the only allowed tool and its been flumoxing people with seemingly random unrelated responses. Often in different languages.
That is when it bothers to respond instead of just sending back an 1099 error code
You just reminded me of how Copilot (also the only allowed tool at some of the orgs I do client work for) will sometimes switch to Toki Pona for no reason.
This is a HUGE clue that someone at Google should probably see…………
Hi, it's Thariq from the Claude Code Team here.
Thanks for the detailed report. We’re confident this is a hallucination but of course take these reports seriously and the team is looking into it. We’ll report back if anything turns up.
I know it's the weekend, so thanks for working hard. Just a suggestion from a user: I wish we could manage Claude Code's memory more easily. Right now, when I go into the .claude folder and change a project folder name or something, sometimes it can't pull up the memory properly. It'd be nice if there were an easier way to import or export it. Thanks!
Piggybacking onto a second/different "just a suggestion from a user":
The VS Code extension needs love. I'm sure you guys are aware but it feels like it is neglected.
GitHub Issues is a graveyard of 3-10 duplicates of really important issues with no activity getting closed. A few examples:
* Lots of /commands missing in between the CLI harness and the VS Code extension
* No way to monitor subagents/tasks/progress visually
* No status bar/line
So the options are this amazing tech is so stupid it just randomly brings up Minecraft or it’s got a major security issue?
The person had “minecraft.py” in their context and the session context was very long.
Having an LLM session with very long context occasionally go off on a tangent is not uncommon. The people who expect absolute perfection out of every LLM interaction see this as some total indictment of the entire technology, but the people who use these tools daily have learned to treat the output as partially stochastic and to avoid extremely long context, even if the model offers it. It’s best to compact strategically or summarize next steps to hand off to a new session. Using sub-sessions can also reduce context pollution at the cost of additional token expenditure to summarize and transfer data to and from the sub-session.
TLDR: the first one
> this amazing tech is so stupid it just randomly brings up Minecraft or it’s got a major security issue
You can sugarcoat it but that's what it is. It's not slightly wrong like a junior engineer or weird like a junior engineer on LSD, it becomes like "your junior engineer suffered a stroke or sudden onset dementia completely forgetting the entire point". one trigger word and that's it we're building Minecraft castles now.
It’s the weekend so we’re allowed to anthropomorphize.
I’ve known some brilliant engineers who would also just randomly bring up Minecraft (more likely Factorio these days) so this makes sense.
¿Por qué no los dos?
What does a self-driving car do when it hallucinates?
Not that different than people, amiright?
---
Note that the author did have a minecraft.py file. So not quite 100% random.
Interesting to see the claudeslop reply as the first comment to the gh post and the reaction to it.
Oh yes, we do not need programmers any more…
it’s the wet dream of execs and pm types. however, i have not seen anything close to it in my life. I remember the UML days, lol. the issue is not the code, it’s the translation layer between business and code. maybe someday ai bridges that gap. history has shown probably not
50% unemployment :D
"Coding is largely solved"
The funny thing is at my current employer, they mentioned that "coding is increasingly becoming a solved problem" and in the same breath, mentioned that one project was too hard for anyone to do so they're not doing it and would rather sell existing features...
Weird. Coding isn’t really “solved” - because “coding” isn’t just the process of typing in characters as fast as possible - BUT the skill floor has been massively lowered while also raising the skill ceiling considerably.
We’re doing projects now that seemed impossible before because we have access to these powerful AI models. They can make things that would have taken weeks or months take days now, freeing up time for even more ambitious buildouts we never would’ve even considered before.
While abused by LLM vendors, that phrase in one form or another I've been hearing since the early '00s and it's likely way older.
Sure but have you ever seen it actually play out in practice like it currently is? Whether or not it's true (of course it's not) people are currently behaving as if it is and firing/hiring accordingly.
Well, when was the last time you wrote machine code by hand?
... but then they went and changed what coding meant.
We've always been layering abstractions on top of abstractions. If we get to an abstraction that works well enough that you no longer have to dive down into the previous layer, we say we've solved coding, and change what coding means. Obviously LLMs aren't there yet.
I love that quote, especially considering the insane amount of bugs that are produced. It’s as easy to debunk as someone claiming ”I can jump to the moon”.
"This thing isn't 100% perfect, contrary to what absolutely no one anywhere said at any time"
In order Fable 5 has rejected:
"Recipe for red-braised pork, I have pork shoulder"
"Write up a framework for MCP patterns I can give to claude code"
"explain the biomechanics of motion in c. elegans" (I get this one, I mostly did it to test and it's related to my hobby project)
Do we get an extra day of functional Fable 5 because it's down?
Not sure the relevance of this comment, but normally if someone built a classifier that bad they’d be fired. Anthropic obviously thinks they have some monopoly power they can use to foist garbage on consumers, I think they don’t.
If people are complaining about Anthropic (on an only-vaguely related thread) rather than simply switching to a suitable competitor, then Anthropic clearly has some 'monopoly' power over the specific capabilities the complainer wants from them.
Fable/Opus 4.8 outperform Codex 5.5 for me at the general architecture/refactoring/performance work I’m doing, to the point where it’s not worth using Codex. Codex will often spit out non idiomatic code that overcomplicates things.
Not to argue the point but that statement isn’t logical, look at all the complaints about restaurants. Publicly complaining about something doesn’t require it be a monopoly.
The consequence of a too strict classifier are annoyed customers who will spend less on Fable. The consequence of a too lax classifier are export restrictions that prevent a huge chunk of their customers from using Fable
I'm annoyed but not surprised at the overeager classification
What does this have to do with anything? Who are you talking to? This is Hacker News, not Anthropic support.
I think they’re just posing the question to other users reading these comments.
HN becoming anthropic support would certainly explain a lot of threads and comments I've seen here lately. Thank you for this.
I'm impressed that folks are using this frontier model for cooking
The safety filter rejected or the model was down?
I asked it how people get blue eyes from their parents and it downgraded me to Opus because of safety.
Is there anything particular about LLMs that would make separating customer data harder than in all SaaS cases?
Yes:
* There's an enormous amount of very expensive shared state (context cache) which you do not want to duplicate when you can avoid it.
* Memory locality is crucially important for performance.
* Hardware is extremely over-subscribed.
* Hardware is extremely expensive.
These factors all make hardware or even traditional memory-space (hypervisor/VM/hardware assisted virtualization) isolation a non-starter for most workloads and customers, which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Moving beyond that, the tools, frameworks, and hardware which the system runs on (GPU) wasn't designed for task isolation and building this isolation is even moreso an emergent research field than it is in x86 CPU hardware-sharing (which has required a huge amount of effort over the past 30+ years to get where we are today).
And, the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
I can't tell if the original post is a real issue or not, but I'm surprised there aren't more like this overall; the whole thing really is a house of cards in this sense.
> which forces all isolation to the software layer. This already makes things way harder than they are in commodity SaaS.
Is this not what happens in most SaaS? Isolation at the software layer? I understand there are special agreements, but they seem to be mostly that – no?
> the ratio of usage/sensitivity to maturity is also just poor overall; these are young companies with rapid development and enormous delivery pressure under incredible customer workload requirements, too.
Mh. The talent density in these companies is apparently quite exceptional. Things like customer data separation is something that is obvious and top of mind. I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
> Is this not what happens in most SaaS?
I think it's fairly popular to try to do more logical isolation in SaaS now, especially with VM-scheduling-as-a-service becoming more popular. For example, I did security architecture at a company who did relatively simple financial processing; we worked to move to a model where customer documents were encrypted using a tenant key which we'd then wrap in both a service key and a login key; users could only get the login key stapled to their session by authenticating against that account, and the processing jobs ran on a cloud vendor's logical isolation. So the user needed a login key, the service needed the attested service key, and the job ran in what amounted to a mini-VM, avoiding issues like "whoops we sent the wrong document ID and the backend gave it back to us" or "whoops, we routed the request to the wrong tenant backend!" This level of isolation would be really hard to achieve in an LLM vendor context.
> I don't see why they would not hire the best to implement these relatively boring/solved things correctly at an architectural level.
I think a lot of these things develop over time; obviously hiring people who have done them before helps, but it's hard. Even the people with strong experience often only know little slices. And unfortunately, every system operating at these scales has emergent behavior which can become really challenging at scale; mistakes like "we used hash(id) as a key in a memory cache without a collision list, and it collided" which would simply never affect most startups become more and more frequent at scale. High rate of change makes it hard to suss these mistakes out and root-cause them, too; "a customer gave us a log where we swapped X and Y" is hard to bisect when you're doing 500 code deploys a day.
Vibe-coding the implementation.
I haven't had much issue with Codex, but seems Claude Code has major issues being reported nearly on the daily.
They also happen to be the most boastful about not reading or looking at the code.
LLMs are very capable, but not nearly to the level they seem to be messaging.
(We've actually moved on from vibe-coding to having the LLM vibe code itself in a loop)
> having the LLM vibe code itself in a loop
The businesslatin name for this is Recursive Self-Improvement
Interestingly I had an almost identical experience to this report in codex. It output a user memory file that looked awfully real and wasn't at all related to my work.
If I had to hazard a guess, doing anything in a multi-tenant way on a GPU is going to be hard mode compared to most SaaS due to lack of memory safe tooling. I've built multi-tenant SaaS systems, and I've done a little GPU programming (a long time ago), but I've never tried to combine the two disciplines.
It'd be terribly compute inefficient to not share prefix caches (KV cache) across customers.
What is the probability that two customers will have exactly the same tokens in cache? Wouldnt it require using the exact same CLAUDE.md, skills, MCPs and context? After that it is even worse since the nondeterminism of LLMs and humans
I suspect what GP is getting at is there will be a strong incentive to implement some structural sharing across tenants to avoid redundantly storing the same tokens over and over. At least I'd be tempted to do this if I was working with a very precious, constrained resource (e.g. VRAM). Doing this correctly seems.. very difficult. [edit] To answer your question directly: the probability that the entire cache is identical between two different users is very low, but the probability that there exists identical chunks of cache between two different users is very high. Exploiting those commonalities successfully will significantly compress the data.
Agree with this and I have been thinking about it recently as well. I think you could implement a cord-like vocabulary to identify large duplicated substrings for exact deduplication and pairwise correlations or vocabulary profiles/small classifiers for forward-looking or speculative deduplications. A clear example is the GPL license, it’s a large substring you might encounter often and highly likely to be accompanied by lots of c code.
This is probably something that you’d be doing on the CPU though before sending anything to the GPU, though that’s definitely the sensitive surface since it’s hardware without good multitenancy. I assume the interface between the CPU and GPU is where you would be most likely to make a mistake where you start decoding data from one fd that was meant for another, or from the wrong position, and get someone else’s data.
I wouldn’t be confident that these are active exploits from deliberately abusing kv cache optimizations though, possibly just the kind of bugs you get from active low level performance tuning/systems work. Since this is something I have seen across providers lately I personally suspect it to be a driver issue.
Given the size of the datacenter class GPUs they're running these models on, don't they need to be processing multiple tenants concurrently per GPU to extract the full potential of the hardware?
I agree, shuffling the data between the CPU and GPU is itself fraught with peril. It's all the hairiest distributed systems problems combined with the sketchiest memory safety issues all in one place.
System prompt for something like Claude Code should be identical, no?
Not only that, but thy need to have the same tokens at the same time? I just can't see how likely this is.
Could just be a bug in the radix tree for the KVCache with deeper, wrong, levels of the trie returning for the same initial prefix match.
Reminds me of a session I had recently (on web!) where claude insisted that i prefixed all my messages with statements about code execution or something, which was not the case. I interrogated it about that and it confirmed that it came from somewhere else, but could not get rid of it and each response mentioned that its gonna ignore those instructions. Eerie.
Anthropic injects text into the conversation triggered by certain conversation topics. This happened to me in relation to some red-teaming related discussion that was adjacent to something “sensitive”, I think sex, and Claude got confused about why I had said some kind of warning and mentioned it it’s response. After a back and forth it was clear that some extra warning to answer but avoid anything inappropriate had been inserted into the conversation.
Claude also sometimes mentions getting messages from classifiers, probably related to auto mode. Amusingly enough, when this happens to a subagent/fork, the orchestrator will call these " hallucinations by the subagent"
Can be malware? Something like https://news.ycombinator.com/item?id=48667495
I am facing a billing/subscription problem and there's nothing I can do or get help on. Their chatbot support shuts me down. Their email is also handled by the chatbot (not even sure whether it's the "same chatbot"). It has been a dead-end. I contacted my bank (credit card issuer) and finally a staffed said I am better off just marking the card lost and having it reissued and that's what I did in the end. I hope that works.
I've never understood in what world this world decided it was okay to hand over these much unchecked power to such corporations. But this is how it has always been one way or the other.
> one tool call result that includes a string that printed a pathname including minecraft.py
This seems like a hallucination.
Could still be a hallucination, but the concerning part is that from the outside a hallucination, local context bleed, and an infra/routing bug can be very hard to distinguish.
Don't worry guys, Anthropic are the experts at security and no one else should have access to bug fixing LLMs because that would be dangerous.
The first reply clearly being a copy and paste from Claude made me want to vomit
If people absolutely need to use AI to write replies, they NEED to start including a "everything after this was generated by AI" disclaimer
Openrouters model providers give me urls people have given them quite frequently.
Note the repro condition: first response after 5+ min, i.e. a cache miss. A cache leak would show up on hits (someone else's cached prefix), not on misses where everything is recomputed from your own tokens.
As a Butlerian, this is hilarious.
Can we acknowledge how it is sad that people get LLMs to basically play computer games for them. What's the point of fun?
happy fourth of july everybody!
Happy fourth to you too :)
fwiw, this could be a bug but the submitters level of arrogance places this rather high on the dunning-kruger side of things. There are multiple other plausible explanations, but this person is probably vibe coder who believes anything an llm says (including explaining its own hallucinations)
Don't worry. Mythos will fix that before release. Oh, wait...
The biggest problem with AI agents is this. You can't debug what the AI is doing, so it's really hard to track down where something went wrong.
What I know for sure:
1.Stuff that has nothing to do with the current session got mixed in.
What guessing:
1.There's a minecraft.py file in the tool folder, and that might have triggered some hallucination.
2.Maybe data from some other project on the user's local machine got mixed in somehow.
3.Or it could be from another user's conversation.
Honestly, if I think about how the system actually works, I don't think it's pulling from another user's data. But other people say they've had issues like that, so I can't completely rule it out.
I saw this thing on YouTube once. When a bunch of users share the same system prompt, or prefix, the computation results get shared through something called a KV Cache. At least, that's what I understood. Not sure if I got it right. But if there's some bug in the hashmap that's supposed to keep those caches separate, then maybe multi-tenant memory management just broke down and that's what caused this. I mean, I can guess, but who knows. And honestly, even if that's exactly what happened, they'd never admit it.
At the end of the day, LLMs are just word predictors, right? They build up some kind of semantic space inside. So maybe the user's question just happened to be near Minecraft in that space. That's kind of what I think.
0 evidence. If this were a real privacy leak, the author would ask their coworker if they talked about the unexpected topic instead of
>"Maybe my coworker was talking about this in another session?"
This would be a critical bug that would slash the market value of a T$ company significantly, go ask your coworker or close the ticket, why do you expect the devs to put an enormous amount of effort hunting a potentially inexistent if you can't make that minuscule debugging effort.
Caching doesn’t work the way the bug reporter implies. Caches are shared (at least across the enterprise), but its key is always a function of the input before it.
We achieved significant savings simply by moving everything that varies across individuals out of the system prompt so every session starts from a cache point.
For example you never want your system prompt to start with the time that the session started. Move that to the first user message if needed.
Caching is not supposed to work like that, but that doesn’t preclude the cache key computation function from having bugs.
Yeah there's quite a lot of potential bugs that could have this shape. If I were to guess it could be a buffer in a buffer pool not being sized and zeroed correctly, allowing stale data to bleed between sessions.
or the cache retrieval function for a key retrieving the wrong entry
There is a massive incentive for optimization, so I expect they’re doing a ton of very clever tricks, all of which make this kind of bug more likely.
Hash functions necesarily have collisions. Also, it is perfectly possible to introduce bugs in the hash function (hash inputs, hash function itself) that allows cross account contamination.
Hash functions necessarily have collisions, but it's perfectly possible to make the expected time between collisions greater than the human lifespan.
There could just also be a bug where the output tokens of session 1 were shared with session 2, due to a race condition or similar.
Seems like a hallucination to me; note that the context contains “unmarkBlock” as the function name, which invites a connection to Minecraft. Still shouldn’t happen of course.
The alternative explanation is that the inference engine, which batches several unrelated requests for parallel processing, messed up the unpacking and returned an unrelated user’s query. This one would be very scary as it will leak arbitrary content, but it seems much less likely here.