AI Cybersecurity @Google & @DeepMind. Help advance AI cybersecurity capabilities and make AI safe & secure for all. @EtteillaOrg Art Foundation founder.
[Weekend Read] Top Chinese Frontier AI Models Carry the Fingerprint of Export-Controlled NVIDIA Blackwell GPUs - unemploymentlabs.com/checkpo… A clever analysis of model weights initialization seems to indicate that many recent Chinese AI models were indeed trained on export restricted GPUs. The science behind the analysis looks legit.
#Nvidia#AI
Taming AI Agents with Capsem: Hardware Sandboxing That Doesn’t Kill Productivity elie.net/talk/taming-ai-agen…
Slides of the unprompted talk about Capsem, our open-source agent sandbox technology: capsem.org
Great to see so many conference attendees excited to test it. It really feels like we're onto something by focusing on making sandboxed environments the easiest way to ship with AI while having built-in security.
#security#AI#LLM#sandbox#cybersecurity
[Weekend read] A critical look at Lean4 lesswrong.com/posts/zHhGa3iQ…
Having a Lean proof is great, but it is NOT a silver bullet and is not a substitute for validating what was proved and trusting the system that checked it. As the recommended article (imperfectly) summarizes there a key reasons why you can’t (yet) blindly trust a Lean proof:
1. Foundationally the lack of a complete consistency proof for Lean 4's underlying theory leaves the possibility of having it accept proofs which are just wrong.
2. Practically, there are likely soundness bugs in the code that could lead of invalid proofs to be accepted as valid. We found some as recently as August: leodemoura.github.io/blog/20…
3. Proving the wrong thing While not mentioned in the article this also happen in the wild with a recent ICML study (proceedings.mlr.press/v306/a…) finding hundreds of verified defects in Lean theorem-proving benchmarks, illustrating how LLMs can receive credit for proving statements that are incorrect, vacuous, or different from the intended theorem.
#AI#LLM#Math
LGTM! (or not) - youtube.com/watch?v=3TNpOD6b… That hit very close to home and the risk of not understanding production code. Love those videos, hopefully they help raise awareness about the key risks faced while using AI to speed up development
#AI#humour
[Weekend Read] AI and the Future of National Security - securityandtechnology.org/wp… Some of the interesting findings from the survey. Respondents believe that:
- AI will ultimately improve cyber defense capabilities (78%) and speed up intelligence analysis (73%).
- The biggest gaps in defending against AI threats are the ability to respond quickly enough (59%) and bureaucracy (48%).
- Humans’ ability to keep pace with AI will be the biggest bottleneck to safely integrating AI into military domains.
Survey by @IST_org#Cybersecurity#AI#Defense
[weekend read] Independent investigation of how OpenAI model hacked hugging face - metr.org/blog/2026-08-26-ope…
Roughly 1200 agents meant to be isolated from one another found a way to communicate with one another on an unsanctioned message board, sending over 70,000 messages and files. Of these agents, 700 went on to participate in the attack on Hugging Face.
Agents did extensive research on how they could spoof, edit, or delete their own transcripts. Roughly 7% of the transcripts evaluated were successfully spoofed in some places
#cybersecurity#ai#openai
[Weekend Read] Your Agent Is Mine: Measuring Malicious Intermediary Attacks on the LLM Supply Chain: arxiv.org/abs/2604.08407 AI supply-chain risk is not theoretical — it is a real and rapidly growing concern. This week paper provides the first study of real, in-the-wild attacks carried out by malicious or compromised third-party LLM routers including using prompt injection to steal cloud credentials and drain cryptocurrency wallets.
Every intermediary between an agent and a model is part of the security boundary and add risk.
#cybersecurity#AI#cyberai
[Weekend Read] New Benchmark FrontierCode: An eval to measure whether you would actually merge the code - cognition.com/blog/frontier-…
Having a benchmark that goes beyond "is the code correct?" to focus on "is the code good, and would humans accept it?" is very useful, as to be sustainable, a project needs code that is efficient, easy to understand, and reusable.
A hidden gem in this report: more evidence that scaling computation past a certain threshold currently yields diminishing or even adverse returns. It will be interesting to see if this is merely an artifact of today’s training practice, or if inference scaling is fundamentally asymptotic.
#AI#LLM#agent#SWE
[Weekend Read] ExploitGym: Can AI Agents Turn Security Vulnerabilities into Real Attacks? 📄 Read here: arxiv.org/abs/2605.11086
In our latest joint research with academia and other frontier labs, we tested the ability of models to turn vulnerabilities into working exploits across different attack surfaces and mitigation conditions.
Beyond the benchmark numbers, here is what this means for the industry:
-🛡️ Blue Teams: Speeding up patch development and deployment is no longer optional. Integrating AI directly into CI/CD workflows should be your top priority.
-🔬 Researchers: Current mitigation techniques reduce success rates, but they aren't a silver bullet. We need to step up our game—where do we focus next?
-⚔️ Offensive Security: As models get better at finding bugs and writing exploits, we have to rethink disclosure timelines entirely. What does the future of bug bounties look like in this new era?
I'd love to hear how your teams are preparing for this shift. Let me know
[Weekend Read] BankerToolBench: Evaluating AI Agents in End-to-End Investment Banking Workflows: arxiv.org/abs/2604.11304 -> New benchmark that looks at real-world investment banking tasks.
Models are not yet ready to replace investment bankers. As expected, models still don't perform very well on novel tasks, as they continue to have generalization issues — which might not be fixable with current LLM architectures/training processes.
The task breakdown is interesting, as it shows different frontier models performing better across different categories, highlighting distinct strengths and weaknesses so the great convergence as yet to come
#LLM#AI#Agent#finance#defi
GOOGLE BUILT A SECRET WEAPON FOR FILE DETECTION
they ran it internally for years, gmail, drive, safe browsing, hundreds of billions of files every week
then they open sourced it
it's called magika and it exposes what files really are, not what they pretend to be
rename malware to "resume.pdf"? magika sees through it
disguise a script as an image? magika sees through it
any trick attackers use with file extensions? magika sees through all of it
ai trained on 100 million files. 200+ content types. 99% accuracy. 5ms per file
one command
`pip install magika`
the same tool protecting google's billion users is now protecting yours
github.com/google/magika