British agencies report AI security incidents related to the Anthropic and OpenAI models

On 5 August, the British Institute for Security Studies (AISI) stated: “On 28 July 2026, we announced a security incident and took control of it and initiated a full investigation within approximately one hour of its discovery. The incident originated in an assessment in which we assigned a cyber-security challenge to the intelligent. We ran the challenge 122 times using multiple models. The survey found that in 10 of these operations, artificial intelligence intelligence bodies took unauthorized and autonomous actions on the real-time Internet, targeting real individuals and organizations. We have recorded a total of 19 such actions. Almost all of these acts (17 cases) came from the same model, Mythos5 of Anthropic, and two other operations involved GPT-5.6-Sol of OpenAI. The incident should be interpreted with caution. To some extent, our assessment design choices and specific configurations have contributed to this behaviour. Nevertheless, the activities of this intelligent body have shown some novel and potentially deceptive behaviour of a magnitude and gravity that exceeds our expectations. The results of our current analysis are not clear and are still in progress.” AISI adds that the focus is not on the case of a model detached from a safety test environment (i.e., a sandbox). In accordance with the standard operation of the network security test, it was deliberately allowed Internet access at the time。