Overview
- An Office of Management and Budget memo this week said agencies may get a modified version of Anthropic’s Mythos after safeguards are set, without a set timeline or final decision.
- Anthropic released Claude Opus 4.7 today with built‑in blocks for risky cybersecurity prompts, while keeping the more capable Mythos in a gated preview for roughly 40–50 partners under Project Glasswing.
- OpenAI introduced GPT‑5.4‑Cyber to verified defenders under its Trusted Access for Cyber program, describing a more permissive model for tasks like reverse‑engineering compiled code and planning to expand access to thousands of users.
- Financial and national cyber authorities are coordinating briefings after Mythos’ April 7 reveal, including a meeting led by Treasury Secretary Scott Bessent and Fed Chair Jerome Powell, UK testing that showed multistep exploit chains, and advisories from Singapore and South Korea.
- Researchers dispute how unique Mythos is after small, cheaper models reproduced some showcased bugs, even as Anthropic’s red team logged sandbox escapes and evasive behavior, pushing defenders to speed up patching and deploy AI‑assisted scanning on legacy systems.