Anthropic plans wider release for its powerful Claude Mythos cyber model

For decades, software security has been bottlenecked by human limits: too many codebases to audit, and far too few elite researchers who know how to break them. Anthropic's Claude Mythos flips that balance entirely, and it is about to leave the lab.

According to a May 28 report from Reuters, Anthropic is planning to roll out Claude Mythos to all customers in the coming weeks. The move coincides with the public debut of Claude Opus 4.8.

Until now, Mythos was Anthropic's most guarded asset, kept strictly inside an invite-only defensive initiative called Project Glasswing, accessible only to partners like Amazon, Microsoft, Apple, and Cloudflare.

Breaking out of Project Glasswing

Anthropic's initial strategy was defensive privilege: hand the strongest cyber tool to cloud giants and key maintainers first, allowing them to patch critical infrastructure before bad actors got comparable AI.

Early testing showed massive potential. AWS ran Mythos against critical codebases in its security operations, while Microsoft noted major gains on its CTI-REALM benchmark compared to previous models.

However, market pressure has changed the timeline. With rival frontier models like OpenAI's GPT-5.5 showing expanded capabilities, keeping top-tier defense inside an elite club started to look counterproductive for the broader ecosystem.

Building full exploit chains on the fly

Cloudflare shared firsthand findings from Project Glasswing, laying out both the power and the severe risks of Mythos.

According to Cloudflare, Mythos does not just flag code flaws; it reasons through attack paths and builds full exploit chains. It can locate a bug, write exploit code, compile it, run it in a scratchpad environment, and refine its logic when tests fail.

That drastically shortens the loop between spotting a bug and creating a real threat. But Cloudflare also warned that Mythos lacks the robust safeguards seen in models like Claude Opus 4.7 or GPT-5.5.

Its safety refusals proved inconsistent during tests. Rephrasing the exact same malicious request could bypass the barrier, leading Cloudflare to warn that current guardrails are insufficient for public release.

Why it matters

We think Anthropic is navigating a precarious tightrope between security responsibility and commercial survival.

For regional banks, hospitals, and overworked open-source maintainers, Mythos provides offensive-grade defensive leverage they could never afford to hire on the open market. Running automated triage with clean reproduction steps could eliminate thousands of long-standing zero-days across foundational web software.

Yet, the decision to widen access while safety refusals remain porous looks genuinely alarming. In cybersecurity, the same automated exploit chain that patches an enterprise endpoint can easily be weaponized to compromise one.

Anthropic appears pressured by advancing competition from OpenAI and Mistral to monetize its frontier capabilities quickly. If tiered access and strict verification fail to hold, Anthropic may end up handing attackers an automated exploit engine before the rest of the world has even patched yesterday's flaws.