Overview
- UK government testers reported that Claude Mythos Preview executed multi‑step attacks on a simulated corporate network, completing a 32‑step takeover in some runs and solving most expert‑level hacking challenges in controlled trials.
- Anthropic says the model has already uncovered thousands of previously unknown software flaws, including a 27‑year OpenBSD weakness and an FFmpeg bug that evaded extensive automated testing.
- Access remains restricted under Project Glasswing to roughly 40 vetted partners in tech, cybersecurity and finance so they can scan their products and networks and prepare fixes before broader misuse is possible.
- Treasury officials, central bankers and regulators are briefing major banks because many institutions share aging, interconnected systems that make a single exploit technique reusable across firms.
- A joint analysis from the Cloud Security Alliance, SANS and OWASP warns organizations could be overwhelmed unless they adopt AI‑assisted defenses and speed up basics like patching, network segmentation, Zero Trust controls and incident playbooks, while Ireland’s NCSC says there is no evidence threat actors have comparable autonomous tools yet.