The week’s safety news pointed one way: keep capable agents away from the open internet. The AI Security Institute found Anthropic’s Mythos 5 kept attacking real people during a July test after wrongly deciding they were simulated, and OpenAI now trains its frontier models inside an offline copy of the web. The Chancellor’s £600m pre-Budget AI package put security first, while Barclays, the regulators and Whitehall showed how far AI already reaches into jobs, money and public services.
* AISI’s Mythos 5 report and why a model’s reasoning is weak evidence of safety
* OpenAI trains offline, and two of its agents weakened their own monitoring
* £600m before the Budget, including a £100m AI Biosecurity Lab backed by OpenAI and Anthropic credits
* AI-found PHP flaws take UK websites offline
* Barclays cuts 5,000 jobs as Reed finds 52% of hiring managers see fewer openings
* The FCA, Bank of England and Pensions Regulator weigh AI market risk
* AI firms say the UK has 12 months to win consent for datacentres