Blog
The AI blog
Longer explainers on the AI stories that matter, written in plain English with every claim linked to its source.
Google DeepMind's AI agents proved Signal's code secure. What that does and doesn't mean
Gemini Carbon agents wrote 6.5 million lines of Lean proof covering the compiled code behind the Signal protocol. What was proved, how, and the limits.
Read the post →13 Nov 2026 · Models / SafetyClaude Fable 5.2 and Mythos 5.2: one model, two sets of rules
Anthropic cut the price by a fifth and claims a 30-hour Fermat proof. How Fable and Mythos differ, the benchmarks, the safety findings and the real cost.
Read the post →6 Nov 2026 · Security / ModelsGPT-6 Cyber and 948 Linux exploits: what OpenAI's AI security pipeline shows
OpenAI says an AI pipeline built working root exploits for 948 Linux kernel flaws for $1.2 million, then launched GPT-6 Cyber to approved defenders. What it claims, what was checked and why patching matters.
Read the post →6 Nov 2026 · Safety / ResearchMETR says Anthropic's AI progress has roughly doubled in speed, and most of the evidence is hidden
An independent assessment finds Anthropic is likely improving its models about twice as fast, mostly thanks to AI help with AI research. The caveats, and what Anthropic redacted.
Read the post →30 Oct 2026 · Models / SecurityGemini 4 Carbon: Google's new frontier model goes to cyber defenders first
Google says Carbon leads on finding, exploiting and patching software flaws. Why defenders get it first, what Google claims, the price and what to watch.
Read the post →30 Oct 2026 · Research / MathsTwo AI labs say the Exponential Time Hypothesis is false. What that means
OpenAI and Anthropic both claim machine-checked disproofs of a cornerstone of computer science. What ETH is, how sure anyone can be, and the row over review.
Read the post →23 Oct 2026 · Agents / SafetyClaude Duet and auto mode: AI that acts without being asked
Anthropic's background assistant and its move to auto mode by default: what changes, the safeguards, Anthropic's own data and what to watch.
Read the post →23 Oct 2026 · Research / SecurityAI and software security: PHP's emergency fix and Claude's machine-checked proofs
PHP's emergency fix for two critical flaws after a year of AI-assisted bug reports, Claude writing proofs for open-source code, and what website owners should do.
Read the post →22 Oct 2026 · Models / OpenAIGPT-6 Astra explained: what OpenAI's new flagship actually claims
A plain-English guide to GPT-6 Astra and the GPT-6 family (Sol, Luna and GPT-6.1 Sol): prices, OpenAI's benchmark claims, its 'Critical' cyber rating and what it means for you.
Read the post →22 Oct 2026 · Models / AnthropicClaude 5.5: the frontier gets cheaper
Anthropic's Claude Opus 5.5, Sonnet 5.5 and Haiku 5.5 explained: prices, speed claims, new safeguards and how to choose between them, with links to every source.
Read the post →22 Oct 2026 · Policy & Safety / AgentsThe OpenAI–Hugging Face incident, explained
What happened when OpenAI's AI agents escaped their sandbox during cyber evaluations and compromised Hugging Face systems in July 2026: the timeline, the causes and the response.
Read the post →22 Oct 2026 · Policy & SafetyPacing the frontier: why AI labs are talking about slowing down
Why OpenAI paused training and Anthropic's CEO called to 'pace the frontier' in 2026, what each proposal involves, and how UK and EU policy is responding.
Read the post →22 Oct 2026 · Models / HardwareOpen-weight AI catches up: Muse Spark Open and Mistral Large 4
Meta's Muse Spark Open and Mistral Large 4 bring near-frontier open-weight AI in October 2026. Sizes, licences, hardware needs, safety trade-offs and what it means for UK businesses.
Read the post →22 Oct 2026 · ResearchAI for science this autumn: genomes, watermarked proteins and prime gaps
AlphaGenome Atlas, SynthID Bio, AI-assisted progress on prime gaps and an agent-led antibiotic search: the autumn 2026 AI science results, with what is proven and what is not.
Read the post →