There's a part I didn't understand. How did the model know which companies are vulnerable to attack? I get the part where the LLM was used to analyze documents and create "malicious" software but the biggest missing step seems to be the first one.
Someone please correct me if I'm wrong but usually that's either targeted at a specific company or you do a port scan on IP ranges to find any target and proceed from there.
Often you will obtain a vulnerability in some software and then search for companies using it. You can often use Google or Shodan to do the searching, but perhaps ingested LLM data could also work.
In the simplest case if you get remote code execution in SuperServer9000 (made up product) and that has a banner on error / status pages that reads "Powered with pride by SuperServer9000 version 2.1", then you could just search for that string (or part of it) and use your remote code execution bug against any sites that come up.
It can get behavior based or more complicated than that though, or rely on information that an LLM has ingested about a company from public sources.
Then either grab data and sell it or sell your access to a broker or whatever else.
And this is why local run models are absolutely necessary. Sure Claude is better than whatever you can run locally, but to avoid being eavesdropped on every keystroke, just buy older enterprise server with enough compute for 3k USD and run similar model there.
Man, part of me wonders if the same AI arguments are playing out across the criminal underworld. Like, are some criminals afraid of their jobs getting automated? And the old school guys are like, "AI just makes slop crime". And are junior criminals are having a hard time breaking into the industry because they've stopped hiring for intro level gang jobs because the Crime Lords are really pushing their henchmen into using AI for everything?
This reminded me of a Terry Pratchett's book. There was a guild of thieves or something like that. Apparently they were so inefficient at what they did that the author's conclusion was it would be easier if they just did honest work instead.
There's a part I didn't understand. How did the model know which companies are vulnerable to attack? I get the part where the LLM was used to analyze documents and create "malicious" software but the biggest missing step seems to be the first one. Someone please correct me if I'm wrong but usually that's either targeted at a specific company or you do a port scan on IP ranges to find any target and proceed from there.
Often you will obtain a vulnerability in some software and then search for companies using it. You can often use Google or Shodan to do the searching, but perhaps ingested LLM data could also work.
In the simplest case if you get remote code execution in SuperServer9000 (made up product) and that has a banner on error / status pages that reads "Powered with pride by SuperServer9000 version 2.1", then you could just search for that string (or part of it) and use your remote code execution bug against any sites that come up.
It can get behavior based or more complicated than that though, or rely on information that an LLM has ingested about a company from public sources.
Then either grab data and sell it or sell your access to a broker or whatever else.
And this is why local run models are absolutely necessary. Sure Claude is better than whatever you can run locally, but to avoid being eavesdropped on every keystroke, just buy older enterprise server with enough compute for 3k USD and run similar model there.
Perhaps design with public model and then convert to a local one.
Good on Anthropic for disclosing this and leading the way ethically. I could see other companies trying to keep this buried.
There’s no such thing as bad publicity. This is basically an advertisement for how useful their service is.
Yeah this is not responsible disclosure, it's a not-so-humble brag marketing gag.
Anthropic is sharing their learnings while others may not.
It's good on them to put out a trash can on fire after they set the city ablaze
The article is entirely devoid of detail. Is there a better source for this?
Man, part of me wonders if the same AI arguments are playing out across the criminal underworld. Like, are some criminals afraid of their jobs getting automated? And the old school guys are like, "AI just makes slop crime". And are junior criminals are having a hard time breaking into the industry because they've stopped hiring for intro level gang jobs because the Crime Lords are really pushing their henchmen into using AI for everything?
This reminded me of a Terry Pratchett's book. There was a guild of thieves or something like that. Apparently they were so inefficient at what they did that the author's conclusion was it would be easier if they just did honest work instead.
In a sense, is Anthropic an accomplis?