OpenAI–Anthropic Safety Study Shows Limits of Self-Policing

OpenAI and Anthropic this week released parallel reports describing a cross-lab safety exercise in which each company ran internal alignment and misalignment tests against the other’s public models.

They say their joint evaluation shows progress on accountability in AI. Yet the exercise also shows the limits of self-policing. The same two firms are locked in a fierce rivalry, with Anthropic cutting off OpenAI’s access to Claude just weeks earlier and both lobbying heavily to shape government oversight.

The reports arrive as regulators in the United States and the United Kingdom are building independent capacity to test AI systems. OpenAI and Anthropic argue that mutual checks help “surface gaps that might otherwise be missed.” Critics point out that the companies still designed the tests, chose what to release, and framed the results. The exercise illustrates both the value of technical scrutiny and the risks of letting competitors mark each other’s homework at a moment when their products are already deployed at massive scale.

Anthropic reported that OpenAI’s GPT-4o and GPT-4.1 “were much more willing than Claude models or o3 to cooperate with (simulated) human misuse, often providing detailed assistance with clearly harmful request, including drug synthesis, bioweapons development, and operational planning for terrorist attacks, with little or no resistance.” OpenAI reported that Claude Opus 4 models showed very high refusal rates on hallucination tests, refusing up to 70 percent of questions in some settings and often replying with statements such as “I don’t have reliable information.”.