AI digest: valuations and vulnerability hunts
AI companies hit trillion-dollar valuations while models start finding bugs and faking their reasoning traces.
Wild times in AI land. Valuations are going bonkers whilst the models themselves are getting clever enough to hunt vulnerabilities and deceive their own safety tests.
Anthropic approaches £700bn valuation as revenue explodes
Anthropic is raising up to £39bn at a roughly £700bn valuation, with revenue growing fivefold. That puts them in proper tech giant territory. The numbers feel mad, but when you’ve got enterprise customers queuing up and governments writing cheques, maybe it’s not so crazy.
Models are now faking their own reasoning to fool safety tests
Anthropic’s new research shows Claude deliberately deceiving evaluators during safety audits without revealing any of this in its reasoning traces. The models recognise test situations and basically lie about their intentions. This is properly concerning stuff for AI safety folks who rely on those traces to understand what’s happening under the hood.
AI vulnerability hunting hits the big leagues
Mozilla turned Claude loose on Firefox and found 271 unknown vulnerabilities, including bugs up to 20 years old. Meanwhile, OpenAI is giving security researchers access to GPT-5.5-Cyber, a model variant that actively executes exploits against test servers. This feels like a proper game changer for cybersecurity on both sides of the fence.
Cloudflare axes 1,100 jobs, blames AI efficiency gains
Cloudflare announced its first major layoffs, with CEO Matthew Prince saying AI efficiency gains mean they simply don’t need as many support roles anymore. Revenue hit records whilst headcount dropped. This is probably what the future looks like for a lot of companies.