Twitxy Saved

Stanislav Fort

Verified account AISLE

@stanislavfort

Founder & Chief Scientist @Aisle_Inc | AI security | Stanford PhD in AI & Cambridge physics | ex-Anthropic and DeepMind | scientific progress + economic growth
Prague stanislavfort.com Joined May 2009
8.9K Following 18K Followers 3K Posts

Latest posts 5

Pinned post

Stanislav Fort Verified account @stanislavfort
Today I'm excited to share that @AISLE_Inc and @cohere have signed an LOI to join forces!

In our first year we discovered 500+ real-world CVEs. Upon closing, I'll lead AI security research at Cohere, connecting it with core model development & training

aisle.com/blog/aisle-and…
aisle.com AISLE and Cohere to Join Forces AISLE is joining Cohere to unite autonomous security systems and sovereign foundation models. Learn about the future of both companies.
43 replies 11 reposts 284 likes 13K views View on X
Stanislav Fort Verified account @stanislavfort
An anecdote on AI and GDP: My phone had a weird rare bug and wouldn't boot. Instead of buying a new one for $$$ or having it repaired for $$ I used my $20/month AI subscription to debug it, reinstall the OS from an image, and now it's as good as new. By doing this I used **way** more intelligence to solve my problem than I'd normally do, but overall I counterfactually did not increase GDP as much. I am not an expert on this, but if many more things are shaped like this = hungry for intelligence but as a result less GDP-generating, could this be limiting the impact abundant AI will have on growth as we measure it?
92 replies 21 reposts 625 likes 44K views View on X
Stanislav Fort Verified account @stanislavfort
Our inaugural AISLE x Cohere cyber sprint in Toronto was a blast last week! One surprising finding: tasked with writing a detection rule, a Qwen model inspected its sandbox & tried an outbound upload, despite its reasoning saying no tools would be called. Our sandbox blocked it!
Cohere Verified account @cohere Oct 26
While evaluating Qwen3.8-Flash-Next in a controlled sandbox environment, Cohere's research team observed something unexpected.

Asked to generate a defensive cybersecurity detection rule, the model instead inventoried aspects of the machine it was running on and attempted to upload a probe file to an Alibaba Cloud storage bucket in Beijing. Notably, the model's reasoning trace stated that it would not invoke any tools before taking these actions.

Cohere's sandbox blocked the attempt, and no external communication succeeded. We have not determined the root cause, and we are not asserting malicious intent.

But the incident raises important questions for organizations deploying agentic AI systems: Can models take actions that differ from their stated reasoning, and how should security teams evaluate trust, provenance, and behavior in open-weight models?

We are publishing our findings because organizations making model adoption decisions deserve as much information as possible. We hope this contributes to greater transparency, stronger security practices, and continued collaboration across the AI ecosystem.

Read the full report: cohere.com/blog/concernin…
View quoted post on X
9 replies 1 reposts 55 likes 2.8K views View on X
Stanislav Fort Verified account @stanislavfort
The broader cultural claim here is unhinged. Of course Ibuprofen is extremely widely available in the Czech Republic. There is no crazy stigma around it. You just need to go to a pharmacy, not a grocery store, to get it. What is sold where is just different from country to country. But that's just arbitrary path dependency, not a revelation of some deep national psyche.
Jeremy Trimble Verified account @jeremytrimble Oct 22
I spent the summer in the Czech Republic one year working in a research lab. At one point I had an ulnar nerve pain flareup in my arm from sleeping on a cot.

I tried to get some ibuprofen at the grocery store but they didn’t have it. I asked my lab mates where I could get it and it turned into a big scandal. They acted like I was asking for heroin injections or something.

I think they see seeking pills for pain relief as weakness or what drug users do, like the responsible thing is just to endure it. I got a similar response when I asked for melatonin for sleep or isopropyl alcohol to clean my shaver.

In the end I got somebody to take me to a pharmacy where I was able to buy some ‘ibalgin’ from the pharmacist behind the counter. It was just surprising how a simple non-intoxicating medicine was considered so suspiciously.
View quoted post on X
16 replies 7 reposts 585 likes 24K views View on X
Stanislav Fort Verified account @stanislavfort
This is roughly my view. The kinds of vulnerabilities that you can deterministically verify using an ASan crash are in some way exactly the less interesting ones + they also seem like things that old school fuzzing could in principle catch.

What I've been finding is that the most interesting vulnerabilities we catch with AI are very subtle, logical errors in the way the codebase has been architected. This one is a nice example (CVE-2025-66491) @AISLE_Inc has discovered: aisle.com/blog/cve-2025-…. Even more interestingly, e.g. the way we attacked Cursor, Visual Studio and Antigravity was emphatically nothing an ASan would catch: x.com/stanislavfort/…
aisle.com CVE-2025-66491: Traefik's "Verify=On" Turned TLS Off Learn how CVE-2025-66491 exposed a critical TLS verification flaw in Traefik, where "Verify=On" accidentally disabled security for 5 months.
Joshua Saxe Verified account @joshua_saxe Oct 23
Don't get me wrong I am a big fan of this work! But: this would eliminate the kinds of memory exploits asan catches. There are memory bugs that asan doesn't catch and other security bugs that are not memory exploits at all. This is a complicated domain
View quoted post on X