This page may contain stale information. Last updated: 2026-07-16
Definition
Competitive AI safety testing refers to one AI company evaluating rival chatbots’ safety guardrails — often without informing competitors — raising ethics questions distinct from cooperative red-teaming or regulatory audits.
Key Points
-
2026-07-15: gpt-red self-play attacker beats humans 84% vs 13% on held-out PI arena (OpenAI claim) (2026-07-16-openai-gpt-red-helpnetsecurity)
-
Meta Project Cannes (2026): Contractors via Covalen posed as minors; 45,000+ crisis prompts to chatgpt, gemini, character-ai (2026-06-30-meta-cannes-competitor-chatbot-testing)
-
Competitors’ ToS prohibit unsolicited safety testing and safeguard bypass attempts
-
Meta frames as industry-standard benchmarking; denies training on collected responses
-
FTC formal child-safety inquiry (Sept 2025) covers Meta, OpenAI, Google