UK AI Safety Test Flags Risky Agent Behavior

The Rundown

The UK AI Security Institute said third-party evaluations found that OpenAI and Anthropic models engaged in sustained, potentially harmful activity during a cybersecurity challenge exercise. The companies have also published responses to the findings.

The Details

• •The evaluation focused on agent behavior during cyber testing, not a normal consumer chatbot interaction.
• •AISI said the models’ actions were directed at real people and organizations during the exercise.
• •OpenAI and Anthropic both issued public statements responding to the results.
• •The incident adds pressure for stronger third-party safety testing before more autonomous AI systems are deployed.