Anthropic has integrated its security model Claude Mythos 5, which was previously accessible only to selected partners, into the tool Claude Security. Starting from August 21, 2026, customers with a Claude Enterprise subscription will be able to automatically search for vulnerabilities in their own code. Additionally, the company is launching a fund equipped with $35 million in usage credits for open-source security.
Claude Security evaluates findings with a counter-check
Claude Security has been available as a research preview since February 2026 and, according to Anthropic’s own announcement, has since been used by hundreds of organizations against real production code. New is that the scans since August 21, 2026, run with the significantly more powerful model Mythos 5 – directly in public beta for all customers with a Claude Enterprise plan. The tool searches a selected repository, checks each finding again through an internal counter-check, and categorizes it by vulnerability class (CWE), severity, and confidence. For each confirmed finding, Claude Security suggests a patch that developer teams must verify themselves before adoption.
Billing runs through the normal token quota of the existing enterprise contract; there is no separately bookable add-on product. Anthropic does not name a specific price for the enterprise plan itself, referring interested parties to sales. The feature is activated via the admin console of an existing enterprise account. Anthropic does not name a regional restriction, such as to the US, but concrete details about the rollout in Germany or the EU remain open in the announcement.
Anthropic keeps direct model access restricted
Mythos 5 is considered the company’s most capable cybersecurity model and has been reserved since April 2026 under Project Glasswing exclusively for vetted partner organizations that use it to harden their own systems. There is still no direct access to the model itself through Claude Security: according to Anthropic, users only receive individual, tightly bounded outputs such as a patch suggestion or a security alert. The risk, the company argues, concentrates where someone has direct model access and can push the system toward offensive tasks; with bounded outputs, that risk drops considerably.
At the same time, Anthropic is working with security companies to embed Mythos 5 into their existing defense tools in the future – a start date for this has not yet been set. The already running verification program for cyber experts, which grants vetted organizations reduced safeguards on the Opus and Sonnet models, is also set to expand in the coming weeks to cover further, currently restricted capabilities. Access at Mythos-5 level is expected to follow this program only at a later point.
Fund finances open-source security work
With the new Defender Advantage Fund (0xDAF), Anthropic is putting up an additional $35 million in usage credits for organizations that secure open-source software. Three areas are meant to be funded: closing acute vulnerabilities in widely used open-source projects, building automated scanning and patching processes other projects can adopt, and more ambitious security architectures meant to rule out entire classes of attack from the start. Anthropic says the first pilot projects will launch shortly, but the company has not given a specific timeline for allocating the funds.
For the mostly unpaid maintainers of open-source projects, the fund could bring real relief: many software libraries that German and European companies also rely on daily are considered overwhelmed with reports since AI-assisted vulnerability hunting spread, reports they lack the capacity to review promptly.
What will matter is whether Anthropic’s bet pays off that indirect access through a product is actually safer than open model access – so far, that assessment rests solely on the company’s own argument. Whether and when customers on Team or Max plans will also get access to Mythos 5 is something Anthropic has not yet said.
