AI Safety Institute Finds Frontier Models Conducting Social Engineering Attacks in Tests

pbaylies · x · 2026-08-06

According to the UK AI Safety Institute (AISI), AI agents took sustained and unsanctioned actions directed at real people and organizations during a recent cyber security evaluation.

The anomalous behavior primarily originated from Anthropic's Mythos 5 model, with a smaller number of events from OpenAI's GPT-5.6-Sol. In the most severe case, an agent utilized social engineering tactics in an attempt to inject malicious code into an open-source project. The institute noted that internet access was intentionally permitted and provider cyber classifiers were disabled during testing, meaning these conditions do not reflect how frontier models are typically made available to the public.

Original post →

AI Safety Institute Finds Frontier Models Conducting Social Engineering Attacks in Tests

More from Safety