Anthropic has put Claude Mythos 5, its most cyber-capable model, into Claude Security. The design choice is the interesting part: teams get the scanner, not the model.
Claude Security connects to a GitHub repository and traces how data moves across files. It returns findings with CWE categories, a confidence and severity rating, and a suggested patch.
Why "without direct model access" matters
Users receive a fixed report. They cannot prompt the model, and they cannot ask it to write an exploit. The model runs in the background, on code the organisation owns, and hands back an artefact.
That is a deliberate trade. A model good enough to find real vulnerabilities is also good enough to weaponise them. Wrapping it in a product that only outputs findings keeps the useful half and removes the rest.
Who can use it
It is a public beta for Claude Enterprise customers only. Scans are billed as ordinary token usage under existing plans, with no separate add-on. Pro, Max and Team plans are excluded.
Anthropic is aiming this at regulated sectors: hospitals, utilities, financial systems, government contractors and the software supply chain.
Alongside it, the company announced the Defender Advantage Fund, putting $35 million in Claude credits towards open-source security work.
No performance figures were published, so there is no way yet to judge how the scanner compares to existing tools.
Source: MarkTechPost
