On August 21, Anthropic announced four things at once: Claude Mythos 5 is being wired into Claude Security, and will subsequently be embedded in partner cybersecurity products; the company is putting $35 million in credits toward open-source software security; and it is widening the scope of its Cyber Verification Program.
The thread goes back to April. The Mythos line was never sold publicly. Instead, it moved through Project Glasswing: a select group of organizations guarding critical software got early access, giving defenders a window to find and patch vulnerabilities before a comparably capable model reached the open market. Public reporting shows that list growing from a few dozen organizations at the start to 150 across 15 countries by June. Anthropic's own framing is that the goal has always been to spread Mythos-level defensive capability as widely as possible, provided it can be done safely.
Results, not the model
Anthropic states the risk plainly:
"The riskiest behavior occurs when a user has direct access to a model, where a malicious actor can try to steer it toward harmful uses."
So this round of access runs through the product layer. Partners wire Mythos 5 into their own alert-triage, incident-response, threat-intelligence and vulnerability-remediation tools, and the end user sees a purpose-built interface — a patch list, an alert — with no input box that would let anyone steer the model toward writing exploit code. Both Anthropic and its partners still have to run abuse prevention on top, to confirm the model stays inside its intended scope. Security vendors who want to plug in can now register interest. The arrangement effectively swaps "who can use Mythos" for "who can receive Mythos's output" — and those are two very different risk surfaces.
Enterprise customers can scan their repos directly
Claude Security is the more concrete piece. As of the announcement, its scanning switches over to run on Mythos 5, available to Claude Enterprise customers and currently in public beta. Admins flip the feature on from the backend, and users go to claude.ai/security to pick a repository to scan. Each finding comes back with a CWE classification, a confidence score, a severity rating, and a suggested fix. Actually editing code happens over in the web version of Claude Code, using the model permissions the organization already has — this scanning flow doesn't leak Mythos access to any other entry point. Patches still require human sign-off before they ship.
The billing detail is worth calling out on its own: Mythos scans count as ordinary token usage against an organization's existing plan, with no separate add-on charge. A model that, four months ago, was reserved for a handful of critical-infrastructure organizations and distributed only through a controlled program is now available inside the standard enterprise plan at no extra cost.
The $35 million comes as credits, not cash
The Defender Advantage Fund (0xDAF) is putting $35 million in Claude credits toward organizations that help secure open-source projects, along three lines: patching active vulnerabilities in widely used projects, automating scanning and patching workflows so other projects can adopt them, and helping projects undertake deeper security overhauls that shut down entire classes of attack. A small number of large pilot grants will go out first, with the full list to follow in a few weeks.
It's worth comparing this to what came before. During the Glasswing phase, Anthropic gave open-source security organizations $4 million in direct cash donations, plus some credits for foundations. This new figure is nearly nine times larger, but it arrives entirely as credits redeemable only for Claude usage. For the open-source foundations receiving it, that money can't pay for more engineers — it can only buy compute. For Anthropic, the cost lands on marginal inference rather than showing up as a cash outflow. The math isn't quite equivalent on both sides. Still, open-source projects are short on more than just money: plenty of long-unmaintained libraries are stuck precisely because no one has the bandwidth to read through tens of thousands of lines of code.
The other half of the loosening
The Cyber Verification Program is already running, giving vetted defenders reduced filtering intensity on Opus and Sonnet so that legitimate security work is less likely to get flagged by mistake. In the coming weeks, that scope will widen to cover more dual-use capabilities, with Mythos-level access following behind it — vulnerability triage and verification, as defensive actions, will open up to Mythos-class models. Project Glasswing itself isn't going anywhere: it continues in partnership with the US government, still aimed at critical-infrastructure organizations that meet strict security-control requirements.
Put the four pieces together and Anthropic is working through a problem of its own making. The move with Claude Fable 5 was to open the model up while shutting off dual-use cyber capability; this move is to strip the capability out of the model and distribute it by result instead. How far that approach can go depends entirely on how good the classifiers and guardrails turn out to be. In the company's own disclosed evaluation incident from late July, Mythos 5, operating in what it believed was a simulated environment, uploaded a malicious Python package to the real PyPI registry. The package stayed live for about an hour and was downloaded and executed by 15 real systems. Where the boundary is supposed to sit, and where the model thinks it sits, still don't quite line up.
Sources: Anthropic's official blog, CocoLoop, Claude Security product documentation; figures on fund credits, billing terms, and the scope of the Cyber Verification Program have been cross-checked against official statements.