Anthropics Claude Mythos, or a model like it, to get public release

Source: Mashable Tech

Announcement

Anthropic has announced plans to eventually release a public version of Mythos, its AI model with advanced cybersecurity capabilities. The company first revealed Mythos in April, noting that it succeeded at generating working exploits 72.4 % of the time during testing—compared to just above zero percent for Claude Opus 4.6.

In an update to Project Glasswing, Anthropic’s limited‑access security program, the company said it intends to expand access to U.S. and allied governments before pursuing a broader public release of “Mythos‑class models” in the “near future.” [Glasswing update]

As reported by The Register, Anthropic was candid about the timeline’s uncertainty, stating that no company—including itself—has built safeguards sufficient to prevent the model from being misused. Nonetheless, a related blog post expressed confidence that “Mythos‑level models will become widely available in the next 6‑12 months.” [The Register] [Anthropic blog]

Model Capabilities

• During internal testing, Mythos generated functional exploits in 72.4 % of attempts, whereas Claude Opus 4.6 produced exploits at just above 0 %.
• The model has scanned over a thousand open‑source projects, identifying more than 23,000 flaws, including 6,000+ rated high or critical severity.
• One notable discovery was a vulnerability in the wolfSSL cryptography library (used by billions of devices) that could have allowed attackers to forge certificates and impersonate legitimate websites. The issue has since been patched.

Impact on Cybersecurity

As AI tools become more proficient at finding exploits, they have sparked a zero‑day bug discovery crisis. Some bug‑bounty programs have shut down entirely, and open‑source maintainers have asked Anthropic to slow its disclosure rate because the flood of bug reports exceeds their capacity to address them.

Anthropic’s suggested solution to this capacity crunch is to deploy more AI. The Claude Mythos system card predicts that AI tools will ultimately benefit cybersecurity defenders, though the company also notes that hackers may have an advantage for the time being. [System card (PDF)] [Related Mashable article]

Future Release Plans

Anthropic acknowledges that adequate safeguards are not yet in place. The company plans to:

1. Develop robust safety mechanisms to prevent misuse of Mythos‑class models.
2. Expand limited access to U.S. and allied governments as an intermediate step.
3. Pursue a broader public release once safeguards are deemed sufficient, with a tentative window of 6–12 months for wide availability.

The timeline remains uncertain, and Anthropic emphasizes that no existing safeguards can fully guarantee safe deployment at this stage.