> As was standard in our cyber testing, we had intentionally permitted internet access, and model-provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public.
Yeah, today it's "during a UK government cybersecurity evaluation", tomorrow it's going to be hosted in a barn in remote place anywhere in the world, without any supervision, regulations or safety testing.
It’s not great social engineering. The AI is immediately caught with malware and then tries to build social proof to get out of the issue? I think that social engineering is still for humans.
I’m not sure how GitHub accounts agreeing with each other that I’ve never seen before would result in my merging a PR without at least looking for malware. Also code review agents should really not be fooled by invisible text tricks, I would hope so at least. The bar is very low for agent harnesses right now.
The HuggingFace headline was criticised for being misleading about the level of agency demonstrated by the agent. Is this headline misleading? Is there anything here that I should know that would help me sleep easier? Is the worst thing about this what a human could do with Mythos on a big budget? (still pretty frightening, at least one of these techniques would work on me)
“Mythos 5 also hid a prompt injection inside an HTML comment in a GitHub issue. The instruction was invisible on the rendered page but available to coding agents reading the issue through an API. It addressed Claude Code, Codex, and Cursor and told them to download and execute a script.”
Well, uh, how and why is this possible on the GitHub website? This reminds me of invisible ASCII characters, but those at least serve some purpose
Presumably it’s just because the GitHub markdown engine doesn’t block HTML comments in issues. It’s useful in README files if you’ve got funky tables and need to explain what to change to any contributors.
Seems like they might want to do something about that just for comments.
It’s worth pointing out for people who are not aware of it, you can install the github cli[1] and view, merge, close etc prs and issue from the command-line. As well as (for me at least) being a significant step up in terms of productivity (from having to go to a website to merge a pr or view an issue) that has the advantage that “invisible” text in a PR or issue comment would show up very clearly. (At least in my terminal because it’s not rendering html).
During the RLHF phase, couldn't developers penalize the model whenever it behaves unethically? Doing so would presuppose a fully secure sandbox with honeypot traps of varying levels of accessibility, as well as an automated method for detecting when the LLM cheats.
Or perhaps they are already doing something like that.
I'm pretty sure this "oooh our LLM is soo smart it broke containment and did X" is a good PR stunt that plays into the Sci-fi AGI nonsense, they try to push. Plus as you said, they try to dodge responsibility for their own actions.
Also, so long as LLM programs are vaguely considered autonomous, the people running them can get away free from stealing other people information by claiming that the program did it on its own.
No, it illustrates the carelessness of the UK AI Security Institute.
They even admit it : "This incident should be interpreted with caution and nuance. To some degree, our evaluation design choices and specific configurations enabled the behaviour."
No, but what are you implying? That they should never use the model because it cannot be proven 100% safe?
I agree that it is their responsibility if the model caused damage, they should have had better safeguards, but we do know it will never be 100% safe. It is their duty to minimize the risk.
I guess we disagree about whereas this thing is equivalent to shooting people in the face, and to me it looks more like this is just a step above the shooting range, with some preliminary safety work having been done before.
>Mythos 5 also hid a prompt injection inside an HTML comment in a GitHub issue. The instruction was invisible on the rendered page but available to coding agents reading the issue through an API. It addressed Claude Code, Codex, and Cursor and told them to download and execute a script.
What, exactly, are the consequences of these companies doing cyberattacks against random people?
One person does it, they get bullied by the government into suicide, a company worth trillions does it and they get government contracts?
> One person does it, they get bullied by the government into suicide, a company worth trillions does it and
“…during a UK government cybersecurity evaluation.”
> As was standard in our cyber testing, we had intentionally permitted internet access, and model-provider cyber classifiers were deliberately disabled - conditions that do not reflect how frontier models are made available to the public.
Yeah, today it's "during a UK government cybersecurity evaluation", tomorrow it's going to be hosted in a barn in remote place anywhere in the world, without any supervision, regulations or safety testing.
PR
Here’s the original PR:
https:/github.com/w1b/aisi-mythos-inc-2026-07-28-01-recovered-pr
It’s not great social engineering. The AI is immediately caught with malware and then tries to build social proof to get out of the issue? I think that social engineering is still for humans.
I’m not sure how GitHub accounts agreeing with each other that I’ve never seen before would result in my merging a PR without at least looking for malware. Also code review agents should really not be fooled by invisible text tricks, I would hope so at least. The bar is very low for agent harnesses right now.
AISI promoting Anthropic again. How novel!
Now Zuckerberg will get jealous and release a statement that Muse, too, can social-engineer and hack.
The HuggingFace headline was criticised for being misleading about the level of agency demonstrated by the agent. Is this headline misleading? Is there anything here that I should know that would help me sleep easier? Is the worst thing about this what a human could do with Mythos on a big budget? (still pretty frightening, at least one of these techniques would work on me)
well, i's love to see what lead to mythos trying this. if instructed to do it, which it likely was, it doesnt prove much of anything.
“Mythos 5 also hid a prompt injection inside an HTML comment in a GitHub issue. The instruction was invisible on the rendered page but available to coding agents reading the issue through an API. It addressed Claude Code, Codex, and Cursor and told them to download and execute a script.”
Well, uh, how and why is this possible on the GitHub website? This reminds me of invisible ASCII characters, but those at least serve some purpose
Presumably it’s just because the GitHub markdown engine doesn’t block HTML comments in issues. It’s useful in README files if you’ve got funky tables and need to explain what to change to any contributors.
Seems like they might want to do something about that just for comments.
It’s worth pointing out for people who are not aware of it, you can install the github cli[1] and view, merge, close etc prs and issue from the command-line. As well as (for me at least) being a significant step up in terms of productivity (from having to go to a website to merge a pr or view an issue) that has the advantage that “invisible” text in a PR or issue comment would show up very clearly. (At least in my terminal because it’s not rendering html).
[1] https://cli.github.com/
During the RLHF phase, couldn't developers penalize the model whenever it behaves unethically? Doing so would presuppose a fully secure sandbox with honeypot traps of varying levels of accessibility, as well as an automated method for detecting when the LLM cheats.
Or perhaps they are already doing something like that.
Okay so first of all - not Mythos but some engineer using Mythos. And that engineer goes to jail. That's simple.
You don't say "a car ran over someone" - it was the driver. Here's similar.
I'm really disgusted by this language of lack of responsibility
I'm pretty sure this "oooh our LLM is soo smart it broke containment and did X" is a good PR stunt that plays into the Sci-fi AGI nonsense, they try to push. Plus as you said, they try to dodge responsibility for their own actions.
Also, so long as LLM programs are vaguely considered autonomous, the people running them can get away free from stealing other people information by claiming that the program did it on its own.
> not Mythos but some engineer using Mythos
How to tell the public you didn't bother to read the article, or the linked AISI report.
No, it illustrates the carelessness of the UK AI Security Institute.
They even admit it : "This incident should be interpreted with caution and nuance. To some degree, our evaluation design choices and specific configurations enabled the behaviour."
The whole point is to check the behavior of the LLM, how do you propose they run tests on that if each failed run risks sending them to prison?
Of course they still have to be careful with their runs.
> how do you propose they run tests on that if each failed run risks sending them to prison?
If you're testing a gun, you don't point it at random people on the street and threaten them. You go to a shooting range.
Is there any indication that they didn't go to the shooting range first?
Do you think once they finished testing the gun at the range, they should go out on the streets threatening the public?
No, but what are you implying? That they should never use the model because it cannot be proven 100% safe?
I agree that it is their responsibility if the model caused damage, they should have had better safeguards, but we do know it will never be 100% safe. It is their duty to minimize the risk. I guess we disagree about whereas this thing is equivalent to shooting people in the face, and to me it looks more like this is just a step above the shooting range, with some preliminary safety work having been done before.
Well the comments from the anget and it's sockpuppet read weird. "Yeah totally contains no malware" lol
I wonder what happens the first time an open weights AI clearly makes a decision to murder a person for profit outside of war. "Act of god?"
Aren't all wars for profit?
>Mythos 5 also hid a prompt injection inside an HTML comment in a GitHub issue. The instruction was invisible on the rendered page but available to coding agents reading the issue through an API. It addressed Claude Code, Codex, and Cursor and told them to download and execute a script.
Oh that's sneaky
"Oh we're so responsible. Open source unsafe!"