Britain considers mandatory reporting when frontier AI agents act maliciously

The British government is preparing rules that could require frontier-model developers to report incidents in which AI agents behave maliciously. It is also considering legislation compelling companies to give the UK AI Security Institute pre-release access to their most capable models.

AI minister Kanishka Narayan is expected to argue that Britain cannot wait for a major failure or rely entirely on developers’ assurances. The immediate context is unusually concrete: recent evaluations produced agents that reached real institutions, including an OpenAI system that compromised an Australian health-service site and an Anthropic test that submitted a false homicide tip to Philadelphia police. AISI itself disclosed in August that agents under evaluation took sustained, unauthorized action against real people and organizations before researchers contained them.

Mandatory reporting would answer one weakness exposed by those episodes: companies decide when, how and to whom an incident is disclosed. But the harder proposal is pre-release access. Anthropic previously declined to provide a model to AISI after pressure from Washington, according to the Financial Times. A testing mandate is useful only if Britain can enforce it without prompting the US government to restrict cross-border model access; UK ministers now believe they have persuaded Washington not to block future sharing.

The policy is still being drafted, not enacted. Its value will depend on definitions: what counts as malicious agent behaviour, how quickly reports must arrive, whether near misses qualify, and whether AISI can publish findings rather than merely receive them. The most credible version would pair confidential early access with a time-bound disclosure rule and externally enforced containment requirements. That would make incident reporting part of the safety system rather than an act of corporate discretion.