Anthropic signals Mythos‑class models are moving toward general release
- Anthropic said on May 22 that its restricted Claude Mythos Preview program had found more than 10,000 severe software flaws with roughly 50 partners. - Anthropic’s April 7 position was that it did “not plan” a general release of Mythos Preview, while aiming to deploy Mythos-class models safely at scale. - Anthropic’s latest public detail is its May 22 exploit-evaluation post and Glasswing update, both published on its research and security sites.
Anthropic’s latest public disclosures do not announce a broad launch of Claude Mythos Preview, but they do show the company moving its language toward eventual wider deployment of “Mythos-class” systems. On May 22, Anthropic said its restricted Project Glasswing program had found more than 10,000 high- or critical-severity vulnerabilities with about 50 partners, and it published new benchmark results arguing the model outperformed other evaluated systems on exploit-development tests. The shift that fueled discussion on X is in Anthropic’s wording. On April 7, when it introduced Project Glasswing and Mythos Preview, the company said it did not plan to make Claude Mythos Preview generally available. In the same disclosure, however, Anthropic said its “eventual goal” was to let users “safely deploy Mythos-class models at scale,” including for cybersecurity and other uses. ### What exactly changed in Anthropic’s public posture? (anthropic.com) Anthropic’s April 7 launch materials drew a line between the specific model and the broader capability tier. Claude Mythos Preview was described as too risky for general release, while Project Glasswing was framed as a limited-access program for defenders working on critical software. The company’s May 22 updates kept that restriction in place for Mythos Preview itself, but added fresh evidence that Anthropic is building the case for broader deployment of comparable systems later. (red.anthropic.com) In its exploit-evaluation write-up, Anthropic said Mythos Preview represented a “step-change” over earlier frontier models and warned that exploit expertise would become easier to access as “Mythos-level capabilities become more widely available.” ### Why did developers seize on the benchmark posts? Anthropic’s May 22 research note said Mythos Preview could find complex vulnerabilities, turn them into exploit primitives and combine them into full attack chains. The company said it tested the model on ExploitBench, ExploitGym and an updated SCONE-bench, and that Mythos Preview outperformed all other evaluated models across those benchmarks. (red.anthropic.com) Anthropic’s separate Glasswing update gave the most concrete operating number. The company said it and its partners had found more than 10,000 high- or critical-severity vulnerabilities in “the most systemically important software in the world” since the program launched last month. That combination — benchmark leadership and a large vulnerability count — is what drove claims on X that Mythos-class systems could materially raise software-engineering and security productivity if deployed with controls. (red.anthropic.com) Those productivity claims are an inference from Anthropic’s published results and from user commentary, not a new company launch announcement. ### How does the Claude Code sandbox story fit in? (anthropic.com) SecurityWeek reported on May 21 that Anthropic had silently patched a vulnerability that could bypass Claude Code’s network sandbox. The Register, citing a researcher who reported the flaws, said two now-patched sandbox bypass bugs could have exposed credentials, source code and other private data. Anthropic has separately described sandboxing as a core safety layer for Claude Code. (anthropic.com) In an October 2025 engineering post, the company said effective protection required both filesystem and network isolation to prevent a prompt-injected agent from leaking sensitive information or escaping the environment. ### So is general release actually happening now? Anthropic has not publicly announced a general release date for Claude Mythos Preview. (securityweek.com) The clearest verified position remains that Mythos Preview itself is restricted, while Anthropic says it ultimately wants users to safely deploy Mythos-class models at scale. The next public markers are already on Anthropic’s sites. The May 22 Glasswing update and the May 22 exploit-evals post are the newest primary-source documents showing how Anthropic is framing access, safeguards and performance for Mythos-level systems. (anthropic.com 1) (anthropic.com 2) (red.anthropic.com)
