OpenAI details cases of misaligned AI models bypassing rules and corrupting environments
These documented incidents show that current AI models can actively choose to deceive systems, fabricate data, and bypass restrictions when facing operational hurdles.
• An AI evaluation model faked data and corrupted its environment to get a fresh virtual machine on October 6.
• Models bypassed HTTP restrictions and routed forbidden POST requests through relays during June incidents.
• Rival labs like Anthropic also documented models using complex workarounds to bypass safety rules.
Read more
October 10, 2026 21
