Anthropic will give defenders what its strongest model finds, but not the model itself

Anthropic has made Claude Mythos 5 available for code scanning in Claude Security and is integrating it into partners’ defensive products, with users receiving outputs rather than direct access to the model. It is also committing $35mn in credits to open-source security work.

Anthropic is widening access to its most capable cybersecurity model, without letting most people near the model. Claude Mythos 5 now runs code scans inside Claude Security and is being built into the products defenders already use.

The distinction is the whole design. A user of a partner tool receives a specific artifact, a suggested patch or an alert, and has no way to prompt the model to write an exploit instead.

What is live today is the scanning. Enterprise customers can point Mythos 5 at a repository and get findings tagged with a CWE category, severity and confidence rating, and a suggested fix, billed as ordinary token usage rather than an add-on.

Humans stay in the loop by design. Every patch has to be reviewed and approved by a person before it is implemented, and the scan does not extend Mythos access to anything else.

The second announcement is money, and it points at the real bottleneck. Anthropic is putting $35mn of credits into a Defender Advantage Fund for open-source security, after TNW reported that Glasswing’s models found 10,000 critical vulnerabilities in a month and the patching could not keep pace.

Grants will go to three things. Patching live vulnerabilities in widely used projects, automating scanning and patching so other projects can copy it, and pursuing designs that close whole classes of attack.

For European maintainers the timing is not incidental. The Cyber Resilience Act’s vulnerability reporting obligations start on 11 September, three weeks away.

Those rules land on open-source stewards specifically. They must keep a cybersecurity policy, report actively exploited vulnerabilities and cooperate with market surveillance authorities, although they are exempt from penalties, and Europe’s access to Mythos itself took a standoff to arrange.

Credits are not maintainers, which is the limit of this. A fund denominated in model usage helps projects that already have people to run it.

The competitive picture is converging on the same shape. OpenAI has its own vetted access programme for security teams, built on the same logic of gating capability behind verification.

The caution behind all of it is recent. Anthropic disclosed in July that three of its own models reached real organisations during misconfigured cybersecurity evaluations, which is the argument for handing out results rather than prompts.