UK unveils AI risk package: wider frontier-AI cyber defence pilots, possible legal reporting of AI incidents and near misses; government confirms AISI is evaluating Gemini 4 Argon

Why it matters

【Key points】

On Oct 12, in his first speech since joining the Cabinet, UK AI Minister Kanishka Narayan announced a package against frontier AI risks: scaling two pilots that use frontier AI for government cyber defence (over 1,000 potential vulnerabilities found and fixed so far); exploring a stronger Cyber Security and Resilience Bill regime covering AI loss of control and reporting of serious incidents and near misses; NCSC and AISI work on responding to incidents involving AI agents; and support for AI verification, including hardware-enabled verification. He confirmed AISI is now evaluating the latest generation of models, naming Gemini 4 Argon, and told labs: if you cannot adequately control the risks, do not release the system.

【Verifiable】

Press release: https://www.gov.uk/government/news/britain-steps-up-national-defences-against-growing-ai-risks

Speech: https://www.gov.uk/government/speeches/building-a-safer-more-secure-britain-in-the-age-of-ai

The release says AISI's evaluations follow work with leading developers and the US government. The speech says the central vulnerability-monitoring service cut time to fix from 50 days to eight, and that Security Minister Dan Jarvis will lead a cross-government cyber programme. Bill details are still to be agreed with Parliament.

【Our read】

1. This officially confirms e/819: AISI is evaluating the newest models, including Gemini 4 Argon. It does not say whether Anthropic's Mythos 5.1 was submitted or whether the US test-first order remains.

2. Only the cyber pilot expansion is concrete; the rest is "will explore". Until incident reporting and loss-of-control rules are law, labs face no new obligation.

3. Singling out AI-agent incidents and near-miss reporting matches the White House's mandatory reporting (e/795), but the UK is going via legislation while the US relies on executive action and voluntary accords (e/815), so obligations on labs may diverge.