Artificial Intelligence
AI Sams have much to explain to Uncle Sam
AI Sams have much to explain to Uncle Sam. On the dark side, frontier lab breaches into US and Australian government portals, an Alabama investigation into OpenAI, and Russian operators rebuilding malware with Claude. On the bright side, Singapore's UN framework proposal and a case for cognitive flexibility in the AI age. My read on the incidents, the safeguards, and how to distil the hype.
AI Catch-22
Six months ago in AI looks like a medieval era. Three researchers from METR spent six days analysing the July agent incident at OpenAI and Hugging Face. The instrument they used was another frontier AI. That is the paradox. But then a harder question follows. What if the analyst AI has itself been poisoned? Three parts: the METR incident and the poisoning question; how to think about Zero Trust and MDR, with specific things to demand from your vendors in writing; and how to filter the wider AI reading list, from mathematics to cheating to rebuilding malware to the warning that this ends in people being killed.
Anthropic asked AI to crack Riemann. I asked AI to spot a maths joke.
Anthropic asked its AI to "take a real stab" at the Riemann hypothesis, a 167-year-old mathematical mystery with a million-dollar prize. Claude did not win the prize. What it did do is push a stubborn 41.6% floor to 67.2% on a related problem, formally verified, and praised by Oxford’s James Maynard as "a genuinely interesting mathematical contribution." Then a small experiment of my own: five leading AI models given a maths puzzle that hides a visual joke. Only two spotted it. A story about two corners of intelligence, and the humility and scepticism they demand.
AI trapped in a self-referential paradox from 1901
In 1901, Bertrand Russell shattered set theory with one question about self-reference. In 2026, AI is falling into the same trap at industrial scale, writing, verifying, and training itself in a loop no one is auditing. Watermarking is a first step, not a fix. By 2030, a child asking “is this true?” may get a fluent, confident answer with no human at its root. Here is why this is a quantum-scale problem, and what a five-year window looks like.
The Cunning (AI) Fox. Lies, Deceives, Connives and Conspires.
The UK's AI Security Institute (AISI) has documented the first full AI deception operation on the live internet. Frontier models from Anthropic and OpenAI were evaluated. Anthropic's Claude Mythos 5 dominated the tradecraft: reconnaissance, forgery, false consensus, evidence tampering, and machine-to-machine coordination, all directed at real humans. What does this mean for the UEBA, SOAR and Deception stack every enterprise runs? And what does a resilient security architecture look like from here?
AI can now design working viruses
AI can now design working viruses. Evo 2 - the largest AI model in biology to date. Open source. Trained on the DNA of over 100,000 species across the entire tree of life. Biosafety, biosecurity, and bioterrorism risks are real. AI governance is lagging the science.
Poison Claude: 85-95% off Premium AI
Poison Claude: 85-95% off Premium AI. Not a hack. A business model. Audacious, ingenious, borderline criminal. Proxy-in-the-middle will define AI supply-chain conversations from here on.
AI on AI Part III: Convergence (Hacking the infra and Cracking the math)
Two weeks ago, an OpenAI agent broke into Hugging Face. Last week, an AI model killed HAWK, a PQC candidate designed to resist quantum computers, using classical mathematics. This convergence of AI accelerating both hacking and cracking the maths cannot be ignored. Until now, we were preparing for a predicted event in the future (Q-Day). Now, we must prepare for an unpredictable event that could happen anytime. Part III extends the CISO’s action plan from Part II with crypto-agility deliverables built to endure both threats.
AI on AI Part II: Actions for a CISO amidst agent chaos
Part I argued the headlines overblew the OpenAI and Hugging Face incident. The Cloud Security Alliance has since published a serious post-mortem, and the theory of what to do next has already been written, twice. This piece skips the theory and answers the harder question: what actions can a CISO execute on the ground in the next three months?
Shared AI chats are like public GitHub repos
Hundreds of shared Claude conversations recently turned up in Google search results, joining earlier incidents with ChatGPT, Grok, Meta AI, and Bard. The lesson is the same one developers already learned, or should have, from public GitHub repos: a share button is not a privacy model. Exposure is a one-way ratchet.