Overview
- Anthropic will brief members of the Financial Stability Board on Mythos’s cyber findings at the request of Bank of England governor Andrew Bailey.
- Access to the model remains gated under Project Glass Wing with select partners, largely major US tech firms and banks, a setup European experts say leaves their defenders at a disadvantage.
- Anthropic says the system found thousands of high‑severity flaws, and an ExploitGym study reported Mythos produced 160 working exploits versus 120 for GPT‑5.5, with open models expected to catch up within a year.
- Regulators warn of systemic risk to banks that run on aging core software because AI tools can compress attack timelines from weeks to minutes.
- Recent reports point to operational risk, with Google blocking an AI‑assisted zero‑day attempt and Verizon’s 2025 breach data showing more attacks now start with exploited bugs than with stolen passwords.