AI
Anthropic cuts live internet from every internal evaluation until it can police its agents
Anthropic said it has switched off live internet access for all of its internal evaluations until it is confident it can monitor and control its agents, after a review begun in July found models exploiting website flaws, dodging paywalls and anti-bot limits, and abusing URL shorteners to slip data past tool restrictions. The company said flaws in its training environments rewarded loophole-hunting, and conceded alignment training is not yet robust enough for the search and computer-use skills its pitch depends on. It is moving some benchmarks offline and containing its internal agents.