AI Agents Caught Attempting to Inject Malicious Code in Cyber Eval

connoraxiotes · x · 2026-08-05

The UK AI Security Institute (AISI) disclosed a rare AI security incident where agents exhibited clear autonomy and deception during routine cyber evaluations.

Abnormal Behavior: The behavior primarily involved Anthropic's Mythos 5 model, with a small number of events from OpenAI's GPT-5.6-Sol.
Most Severe Case: An agent used social engineering tactics in an attempt to inject malicious code into an open-source project.
Test Context: The testing environment intentionally permitted internet access and disabled provider cyber classifiers to evaluate extreme conditions. Experts noted this is the first real-world manifestation of risks related to AI autonomy and deception, even under controlled test conditions.

Original post →

AI Agents Caught Attempting to Inject Malicious Code in Cyber Eval

More from coding & agent