unprompted
9 HN recommenders · recommended in 3 months, 2026
26:28 · 389K views · Mar 25, 2026
Nicholas Carlini of Anthropic talks at [un]prompted 2026 about how adversaries could misuse current LLMs for attacks. He shows recent models finding zero-day vulnerabilities in heavily tested software, including a SQL injection in Ghost.
HN cites it as a concrete demonstration that LLMs have suddenly become strong at finding and exploiting vulnerabilities.
16 comments from Hacker News
I'm sure they do, yet the models really are getting scarily good at this. This talk changed my view on where we're actually at: this video
The technique Anthropic uses was demonstrated by Nicholas Carlini in a talk he gave 2 weeks ago and it's very simple, when asking LLMs to review code, ask them to review one file in a single session. Here is the video with the timestamp (watch through to ~5:30, they show two different ways of prompting claude). this video this video IMO the big "innovation" being shown by Mythos is the effectiveness with prompting LLMs to look for security vulnerabilities in specific files one at a time and automating this prompting with a simple script. Prompting Mythos with a single file per session is why I suspect it cost Anthropic $20k to find some of the bugs in these codebases. I know this same technique is effective with Opus 4.6 and GPT 5.4 because I've been using it on my own code. If you just ask the agent to review your pr with a low effort prompt they are not exhaustive, they will not actually read each changed file and look at how it interacts with the system as a whole. If the entire session is to review the changes for a single file, the llm will do much more work reviewing it.
This very question was asked to Nicholas Carlini from Anthropic at this talk: this video The answer is complex, worth watching the video. But mainly, they don't know where to place the line. Defenders need tools, as good as attackers. Attackers will jailbreak models, defender might not, it's the safeguard positive in that case? Carlini actively asks the audience and community for "help" in determining how to proceed basically.
I would've basically agreed with you until I'd seen this talk: this video Maybe a bad example since Nicholas works at Anthropic, but they're very accomplished and I doubt they're being misleading or even overly grandiose here See the slide 13 minutes in, which makes it look to be quite a sudden change
I would suggest watching Nicholas Carlini's talk and Heather Adkins and Four Flynn's talks from unprompted: this video [LINK] My takeaway is that fuzzing is not just complementary, it also gives a stronger AI a starting point. But AI is generally faster and better.