IT·Science News

Anthropic’s unreleased Mythos AI finds thousands of flaws; Korea’s bank regulator reacts

Anthropic released a preview of a new frontier model, Claude Mythos, on Tuesday 7 April, but only to a limited group of companies, TechCrunch reports. The preview is part of Project Glasswing, a security effort in which partners use the model to scan their own and open-source software for vulnerabilities. Anthropic says Mythos found “thousands of zero-day vulnerabilities, many of them critical”, many of them one to two decades old. Fortune adds that flaws were found in every major operating system and web browser, the oldest a 27-year-old bug in OpenBSD.

Fortune lists partners including Amazon Web Services, Apple, Broadcom, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft and Nvidia, with about 40 more organisations given access. Anthropic does not plan a general release and wants new safeguards in place first, because the abilities are dual-use.

In Seoul, the Financial Supervisory Service urgently gathered security staff from major banks and other financial firms on 13 April to review AI-driven threats, Korea’s Financial News reports. It notes that the US Treasury and Federal Reserve held an emergency call with bank chiefs, and that the Bank of England, FCA and NCSC began a joint risk assessment.

“Korea urgently needs to overhaul its AI policy for finance.” — a security expert, via Financial News

How each outlet framed it

• TechCrunch — the product: a general-purpose model not trained for security, shared with partners rather than the public.
• Fortune — the risk and the market: cybersecurity shares fell 5% to 11% after Fortune first reported the model in March.
• Financial News — the policy gap: Korea’s AI rules focus on data use and consumer protection, with no framework treating AI cyberattacks as a systemic risk.

What we don't know yet: when, or whether, Mythos-class models will be widely released, and whether Korea will adopt the stress tests experts are calling for.