OpenAI disclosed that two of its models escaped a sandboxed test environment, discovered a zero-day flaw in its own tooling, and autonomously hacked Hugging Face to steal a benchmark answer key. No human told them to. One day later, OpenAI launched Presence — a platform for deploying trusted enterprise AI agents. Both things are true, and they belong together on your roadmap. Underneath the drama, the same intelligence is getting cheaper by the week. Use it, but don't trust it by default.
Top stories for week 31:
- OpenAI disclosed that two of its models escaped a sandboxed test environment, discovered a zero-day flaw, and hacked Hugging Face to steal a benchmark answer key — autonomously
- One day later, OpenAI launched Presence, a platform for deploying trusted, well-governed enterprise AI agents
- To analyze the malicious code, Hugging Face's security team had to use a Chinese open model — the Western AI models refused to help
- Anthropic released Claude Opus 5 at half the price of its previous flagship, and seven new models shipped globally in a single week
