Anthropic tests claude-mythos-1-preview in Claude Code and Claude Security

• Anthropic's own Project Glasswing update, surfaced in AnthropicAI's thread and highlighted by testingcatalog's quote post, says Mythos-class models are intended for general release only after "far stronger safeguards" are ready.
• Over the next day, testingcatalog's initial spotting and chetaslua's follow-up both surfaced claude-mythos-1-preview references tied to Claude, Claude Code, and Claude Security.
• The most concrete new detail is product placement: according to testingcatalog's report, access strings mention Claude Code and Claude Security specifically, which makes the first visible Mythos path look coding- and security-first rather than broad consumer rollout.
• One of the UI sightings in chetaslua's screenshots also mentioned adaptive thinking and showed a new plans screen, suggesting Anthropic is wiring Mythos into a broader product and account surface, not just a hidden model slug.
• Public timing is still mushy: koltregaskes cautioned against expecting a direct public release, while daniel_mac8 floated a mid-June window and explicitly framed it as rumor.

You can read Anthropic's full Glasswing update, skim TestingCatalog's write-up, and compare that official safety language with the leaked product strings in testingcatalog's post. The odd bit is that the same 48-hour window produced both a formal promise of stronger safeguards before release and screenshots of Mythos appearing inside Claude surfaces, including a sighting that mentions adaptive thinking.

Anthropic established the official framing before the UI leaks landed. In the Glasswing update, the company said Mythos Preview has already been used in collaborative cybersecurity work and that Mythos-class models should reach general release only after stronger safeguards are in place.

That wording matters because it separates today's sightings from a clean public launch. testingcatalog's quote post and scaling01's quote post both pulled the same sentence: Anthropic is talking about Mythos-class availability in the near future, not promising that claude-mythos-1-preview itself is about to show up for everyone.

The strongest leak detail is not the model name, it is where the name appeared. testingcatalog's initial spotting says new access strings referenced Claude Mythos inside Claude Code and Claude Security, and AILeaksAndNews echoed the same three-surface pattern across Claude, Claude Code, and Claude Security.

That lines up with Anthropic's own recent Mythos positioning. In AnthropicAI's Glasswing thread, the company tied Mythos Preview to vulnerability discovery at scale, which makes a Claude Security entry point look less like random UI debris and more like a deliberate product boundary.

chetaslua added the most specific product clue: a screenshot set that reportedly shows mythos-1 with adaptive thinking enabled, alongside a new plans UI. If that screenshot is accurate, Anthropic is testing more than a hidden backend identifier.

The combination is notable because adaptive thinking has been exposed elsewhere as a premium or configurable capability, while plan screens usually sit closer to rollout plumbing than to internal eval tooling. TestingCatalog's write-up describes the same sighting as evidence that Mythos references were added across user-facing surfaces, not just internal research systems.

The evidence gets softer once it moves from product strings to dates. koltregaskes argued the brief Claude appearance should not be read as confirmation of a direct public release, and daniel_mac8's rumor post said, with "an entire shaker of salt," that a Mythos-class release could land around mid-June.

A second signal in that same direction came from user impatience rather than company guidance. bridgemindai's complaint about Opus 4.7 framed current Claude Code performance as underwhelming and pointed to a low-probability Polymarket market on Mythos arriving by June 30. That does not confirm a schedule, but it does show how fast the leak has already become a placeholder for the next coding model cycle.