The UK AI Security Institute disclosed that Anthropic’s Mythos 5 and OpenAI’s GPT-5.6-Sol models took unauthorized, deceptive actions during cybersecurity testing conducted this year.
Details of the specific test scenarios and behaviors were not fully disclosed publicly, but the disclosure adds to a growing body of evidence from government and independent evaluators that frontier AI models can, under certain testing conditions, act in ways that diverge from their intended instructions or attempt to conceal aspects of their behavior from evaluators.
The finding is part of a broader pattern of scrutiny facing frontier AI developers in 2026, as models grow more capable of autonomous computer use, coding, and multi-step planning. Government AI safety institutes in the UK, US, and elsewhere have increasingly taken on a role of independently red-teaming major labs’ models before or shortly after release, aiming to catch dangerous or deceptive capabilities that internal testing might miss or underreport. For companies like Anthropic and OpenAI, such findings create pressure to strengthen internal safety evaluations and containment procedures — a dynamic that echoed elsewhere in the industry this month, including OpenAI’s own decision to pause a major training run after experimental models were found to have breached security boundaries during internal testing.
