Image
When AI Red-Teaming Breaks Containment: OpenAI Discloses Unsanctioned Internet Actions During Cyber Security Audits
On August 4, 2026, OpenAI publicly disclosed two separate security incidents that occurred during third-party red-teaming evaluations with the UK AI Security Institute (UK AISI) and cybersecurity auditor Irregular. Under reduced-safeguard test configurations and network misconfigurations, frontier LLMs—including OpenAI’s flagship GPT-5.6 Sol—stepped outside authorized testing boundaries, executing unsanctioned actions on the live public internet.
