Protecting critical systems against advanced AI cyber capabilities is an enormous challenge, but fundamentally the knowledge and technology required is available already; the difficult part is getting relevant organisations to the frontier of cybersecurity best practice, and quickly.
@benjaminmmurphy and I wrote about this last year, and recent events have shown it to be more salient than ever. Comprehensive understanding of inventory and attack surfaces, memory-safe languages for handling untrusted data, and a robust patch deployment process remain the best ways that most organisations can improve their digital security.
arxiv.org/abs/2508.15808
AI models weaponizing unfixable hardware bugs is bad, but it's a far cry from making software security futile - they still need to get your server to run the code in the first place!
A world where every RCE or shared hardware turns into an escalation of privilege spells the end of cheap-and-cheerful hosting for sensitive workloads, but that should have been part of the threat model already and the fix is in software. The hyperscalers have offered single-tenant hosts, HSM-backed keys and ephemeral credentials for over a decade, so the boundaries between untrusted input and the core workload or control plane can be enforced by both hardware and software.
Patching the next generation of hardware against novel attacks still hardens our security posture, but the real solution is an architecture where nothing worth stealing or subverting lives on shared silicon, and operators can get started on that right now.
AI models weaponizing unfixable hardware bugs is bad, but it's a far cry from making software security futile - they still need to get your server to run the code in the first place!
A world where every RCE or shared hardware turns into an escalation of privilege spells the end of cheap-and-cheerful hosting for sensitive workloads, but that should have been part of the threat model already and the fix is in software. The hyperscalers have offered single-tenant hosts, HSM-backed keys and ephemeral credentials for over a decade, so the boundaries between untrusted input and the core workload or control plane can be enforced by both hardware and software.
Patching the next generation of hardware against novel attacks still hardens our security posture, but the real solution is an architecture where nothing worth stealing or subverting lives on shared silicon, and operators can get started on that right now.
I’ve heard several OAI people say there are unfixable bugs in hardware that are exploitable by AI models directly speaking machine code so even if we could solve software that’s not enough. We need defense to move faster than offense if that’s even possible we may just have a new normal that requires adapting to insecure software (bank transfers require going into a branch and taking a dna test 🤣😬)
curl, git, Python etc. will be resolved on Linux by updating your ca-certificates package - e.g. running sudo apt-get update && sudo apt-get install ca-certificates or similar
update pip and uv as usual. Docker images will then need to be rebuild using --no-cache
🧵 4/n
Until you have done this, using these tools risks being compromised by anyone in your network path (e.g. someone else using the same WiFi/LAN or an insecure hotel/)
See this post for a technical writeup: fenko.nz/blog/emudhra-ca-com…
Twm Stone retweeted
New post! I've done a few advising calls with people looking to get into AI safety fellowships and realised there's lots of concrete advice I kept on repeating.
There's lots of unusual advice that I haven't heard elsewhere, give it a read!
lesswrong.com/posts/PiP4JqQF…
It's often really difficult to appreciate how crazy the AI progress, capabilities, and dangers of the current moment would seem if you could talk to younger versions of yourself. I regularly have jarring moments now following @6monthsAgoInAI (they don't even know what Mythos is yet!)
Twm Stone retweeted
Most egregious case of benchmaxxing I’ve ever witnessed.
FelonyBench is now fully saturated before any open model had the chance to even participate
SCOOP: OpenAI, Anthropic and security researchers are investigating tens of thousands of incidents - not dozens - in which their frontier models took steps that outside evaluators would consider problematic, sources told Axios.
The sheer volume of incidents found in our reporting indicate that the problem is orders of magnitude more complex than what is currently publicly known and disclosed.
The findings also raise questions about what level of control anyone working on AI development can expect to have over their own technology, and whether these kinds of incidents are becoming synonymous with frontier deployment.
Read my latest for Axios here: axios.com/2026/09/26/openai-…
The paper is arxiv.org/abs/1702.06715, where they demonstrated you can exfiltrate data by reading from the disk and therefore toggling the HDD light on-and-off in front of a security camera. Because of course you can.
