Claude Security
The Strongest Model,
Kept Out of Reach — Mythos 5's Design
Anthropic's vulnerability scanner "Claude Security" now runs on Claude Mythos 5, the company's most capable model. But you can't send Mythos 5 a free-form prompt — all a user ever receives is the scan result, a deliberately different access model from anything Anthropic has shipped before.
What Happened
Putting the Flagship
Model to Work Scanning
Anthropic announced that its vulnerability scanner, Claude Security, now runs on Claude Mythos 5, the flagship model it unveiled in June alongside its sibling, Claude Fable 5. Per Anthropic's official announcement, Claude Security scans codebases, rates vulnerabilities by severity with CWE classifications, and proposes patches — now with Mythos 5 doing the underlying analysis.
The key detail is how it's delivered. According to reporting from MarkTechPost, Mythos 5 is capable enough to simulate attack techniques, but individuals cannot send it free-form instructions. Rather than handing users a prompt box, Claude Security returns only the result of a specific vulnerability-analysis task.
Mythos 5's access terms
The danger was never the model's capability —
it was direct access itself.
Why It Matters
Withholding access
becomes the new safeguard
Anthropic has staked out a clear position: direct access to a model is the riskiest thing you can grant a user.
Mythos 5 launched this June alongside Fable 5, initially restricted to vetted cybersecurity partners through the invite-only "Project Glasswing" program (the US Commerce Department has since cleared broader use for certain trusted partners). Folding it into Claude Security extends that same logic. Summarizing the reasoning quoted in that coverage: when a user has direct access to a model, a malicious actor has the clearest path to steering it toward harmful uses; if users can only receive a specific output — a patch, an alert — that risk drops sharply. In other words, the real safety lever isn't the model's raw capability, but who gets to touch it and how.
Who It Affects
Who feels this, and how
Security engineers
You get flagship-tier vulnerability detection through Claude Security, but you can't repurpose Mythos 5 itself for custom tasks — usage stays confined to the fixed scan-and-patch output.
Risk and leadership teams
A design that only ever hands back a fixed output is easy to justify to auditors, which should ease internal rollout. Just remember direct access to Mythos 5 itself is still limited to a small set of authorized partners.
Anyone building prompt-injection defenses
"Never let users hit the model directly — hand back structured output instead" is a pattern worth stealing for any other high-privilege AI tool you're designing.
What's Next
What to watch for
Track Claude Security's rollout
Availability still looks limited — watch Anthropic's announcements for when your organization becomes eligible.
Watch for wider direct access to Mythos 5
The US Commerce Department has already cleared broader use for trusted partners, so the invite list may keep expanding.
Consider the same pattern for your own high-privilege AI tools
"Structured output only, never a raw prompt" is a constraint worth applying to internal high-risk AI features too.
Risks & Open Questions
Reasons to stay cautious
The design makes sense, but it has limits. First, "structured output only" also means less flexibility — users get scan results and fixed patch suggestions, with little room to probe further when something falls outside known vulnerability patterns. Second, direct access to Mythos 5 itself remains restricted to a small set of invited partners, so for most companies this is a benefit received indirectly rather than a tool they can use themselves. Third, output-restricted access models like this improve safety but can also make a model's decision process more of a black box, raising the risk that users over-trust the scan result without understanding how it was reached.