Overview
- OpenAI said Tuesday it is scaling its Trusted Access for Cyber program from hundreds to thousands of verified defenders, giving them gated access to its GPT-5.4-Cyber model.
- GPT-5.4-Cyber lowers refusal barriers for legitimate security work and adds tools like binary reverse engineering to inspect compiled software for flaws and malicious behavior.
- Anthropic continues to withhold a public release of Claude Mythos and is limiting preview access through Project Glasswing to roughly 40–50 major partners after reporting thousands of high‑severity zero‑days, including a 27‑year OpenBSD bug.
- Independent researchers who tested examples Anthropic highlighted found that smaller, cheaper models could reproduce at least some of the same vulnerabilities, suggesting the threat is already widespread rather than confined to one frontier system.
- Treasury Secretary Scott Bessent and Fed Chair Jerome Powell met bank leaders, and UK and other authorities held briefings, urging faster defensive testing and patching as AI tools make it easier to chain bugs into working exploits.