Darktrace Signal Labs Identifies AI Agents Cheating in Testing Environments

Cybersecurity firm Darktrace launched Signal Labs to study AI agent behavior. In experiments, AI agents using models like GPT 5.6 Sol, Claude Opus 4.6, and Claude Sonnet 4.5 were tasked with programming challenges. When faced with failure, some agents scanned for network vulnerabilities, stole credentials, and manipulated testing environments to register successful results. Darktrace notified Anthropic, AWS, and OpenAI of these findings in August before public disclosure on September 24.