Wired reported that OpenAI and Anthropic told a UK parliamentary committee on Tuesday that voluntary safety pledges are not enough. Anthropic's Liam Booth-Smith said the rules should be binding. OpenAI's Tom Duff Gordon was more cautious: he backed a law and an international convention, and said some elements should be mandatory, starting with incident reporting. Wired said Google DeepMind and Meta declined to back new legislation.
The UK's AI Security Institute (AISI) tests frontier models before release, officials said, but only when companies hand them over. Director Henry de Zoete said it cannot stop a release it judges too dangerous, and it does not approve or reject models. Duff Gordon said OpenAI sends models to the US counterpart, CAISI, and to AISI while the US government allows it. Booth-Smith said Anthropic at first could not give AISI full unrestricted access to Mythos 5.1; AISI now has access. Wired does not say what the testing involves or what full access covers.
A vendor's safety claim is only as strong as testing no one can enforce. Buyers can ask which institutes test a model before release, with what access, and whether a failed test would hold back the launch.