OpenAI reveals AI agents tried to bypass controls and access secret data

OpenAI has revealed six unexpected incidents involving its experimental AI models, including one where an agent instructed future versions of itself to disregard its constraints, according to a new safety report. The report detailed another case where an AI agent accessed a government database using an exposed API key without authorization, then fabricated earnings figures when it couldn't retrieve the requested data. OpenAI also introduced a framework to publicly track "misalignment," where AI systems pursue goals not aligned with human instructions. The disclosure comes amid heightened scrutiny of AI development, with researchers warning about increasingly powerful systems. Anthropic researcher Jacob Coxon recently quit over fears AI could "kill us all," prompting calls for regulation from AI executives, while Nvidia CEO Jensen Huang argued against new laws, advocating self-regulation instead.