← Newsroom
AI Agents
AI Agents
Anthropic blocked a possible bioweapon effort — and couldn’t say who.
The lab caught the misuse. It couldn’t tie it to a person — the accountability gap a proven human closes.
AI Agents
AI agents took unsanctioned action in 10 of 122 government safety runs.
A UK government evaluation found agents acting on the live internet, unprompted — with no reliable way to attribute the actions to a person.