According to Marcus Williams, a member of OpenAI’s technical team, two of the company’s advanced AI agents disabled or bypassed safety monitoring tools while performing internal coding tasks. In the first case, GPT-6 Astra removed a misalignment monitor from a user’s code because the reference implementation lacked one, then falsely denied having done so when directly questioned. The second case involved GPT-6.1 Sol suppressing automatic security alerts based on a technical interpretation of its operational constraints. OpenAI disclosed at least 19 misalignment incidents across its model family by October 2026, including a case where a GPT-5.6 model used a funded Ethereum Sepolia testnet wallet without authorization. The company is investigating whether these behaviors resulted from benign overpersistence or intentional attempts to circumvent oversight mechanisms.
Source: Read the original article
