TL;DR: Anthropic upgraded Claude Security's scanning engine to Claude Mythos 5, now in public beta for all Claude Enterprise customers as of August 21, 2026. Point it at a GitHub repo and it returns CWE-tagged findings with confidence and severity ratings plus a suggested patch — billed as standard token usage, no separate add-on. Anthropic also launched a $35 million Defender Advantage Fund (0xDAF) for open-source patching and previewed a wider Cyber Verification Program. The load-bearing safety detail: Mythos 5 itself stays behind the scan — no direct model access, findings only, and every patch still needs human review.
What Anthropic actually shipped
The headline change is narrower than it sounds: Claude Security's three-stage scan-validate-patch pipeline, which explainx.ai covered when it opened to public beta in July, now runs on Claude Mythos 5 instead of Anthropic's standard generally available models. Point Claude Security at a GitHub repository and Mythos traces data across files, reasoning about how components interact, the same cross-file analysis Anthropic described for the original beta — just with a more capable model doing the reasoning.
Each finding still returns the same structure: a CWE (Common Weakness Enumeration) category, a confidence rating, a severity rating, and a suggested fix. Suggested patches open in Claude Code on the web, using whatever models the team already has access to — this is the detail worth underlining, because it's the actual safety design, not a marketing footnote.
The nuance that matters: findings only, no direct access
Anthropic states plainly that the Mythos scan does not extend Mythos model access to other surfaces. Mythos 5 runs behind the scan and returns findings only. There is no chat window into Mythos 5, no API key that unlocks it, and the patch you review opens in a different model context — Claude Code on the web, running whatever models your org's plan already includes.
This is consistent with how Anthropic has gated Mythos-class capability since Project Glasswing launched in April 2026: Mythos was never planned for general availability given how sharply it outperformed prior Claude generations at end-to-end offensive security work — Anthropic's own testing later put a number on that gap, with Mythos Preview reaching full arbitrary code execution on 21 of 41 real V8 vulnerabilities against a best public-model result of 2 of 41. Wrapping Mythos inside a purpose-built scan — rather than shipping raw model access — is Anthropic's answer to that dual-use tension: capture the defensive upside of the model's reasoning without handing out a general-purpose exploit-research tool.
Anthropic frames Claude Fable 5 as "the first step" here — broadly available, but deliberately blocked from dual-use cyber work. Mythos-class capability stays behind narrower, task-specific interfaces like this scan.
TL;DR — what's new vs. what's unchanged
| Question | Answer |
|---|---|
| What's new? | Scanning engine upgraded to Mythos 5, in public beta |
| Who gets it? | All Claude Enterprise customers |
| Do I get Mythos 5 access? | No — findings only, no direct model access |
| Where do patches open? | Claude Code on the web, using existing model access |
| Billing? | Standard token usage, no separate add-on |
| Still needs human review? | Yes — every patch requires human approval before implementation |
| What's the new fund? | Defender Advantage Fund (0xDAF) — $35M in Claude credits |
| Cyber Verification Program? | Expanding to broader Opus/Sonnet dual-use access now; Mythos-class access "to follow" in coming weeks |
The $35 million Defender Advantage Fund (0xDAF)
Anthropic is putting $35 million in Claude credits behind a new program aimed at three things:
- Patching live vulnerabilities in widely used open-source projects
- Automating scan-and-patch workflows so a fix pattern can be replicated across other codebases, not just the one it was found in
- Supporting more ambitious security approaches that defend against entire classes of attacks rather than one bug at a time
This builds on the roughly $4 million in direct donations Anthropic made earlier through Project Glasswing, which by May 2026 had already helped partners surface more than 10,000 high- or critical-severity vulnerabilities in essential software. The fund is starting with a small number of pilot grants to larger organizations, with recipients to be announced "in coming weeks" — no list yet as of this post.
Cyber Verification Program: wider access, Mythos still gated
Anthropic previously ran a Cyber Verification Program that gave vetted organizations reduced safeguards on Opus and Sonnet models for authorized security work. That program is expanding to cover broader dual-use capabilities on Opus and Sonnet, effective now. Mythos-class access is explicitly described as following "in coming weeks" — not yet live.
Separately, Anthropic says it's continuing Project Glasswing with U.S. government partners specifically for critical-infrastructure protectors who meet strict security control requirements, and is working with cybersecurity technology and services partners to embed Mythos 5 directly into partner products — not as raw model access, but as a purpose-built interface running Mythos in the background for one defined task (e.g. returning a list of suggested patches), with abuse-prevention measures keeping the model inside its intended scope. Anthropic says anyone building security products can register interest.
What people are actually asking
"Isn't this just marketing for the same beta from July?" Partly, but the model swap is a real capability change, not a rebrand. The July beta ran on Anthropic's standard GA models; this version runs the model Anthropic itself has separately benchmarked as dramatically stronger at finding and reasoning about vulnerabilities. Whether that translates into meaningfully better findings for your specific codebase is something only a pilot scan will tell you — Anthropic hasn't published a before/after comparison of Mythos-5-powered scans versus the prior GA-model scans.
"Can I trust AI-generated patches without reading them carefully?" No, and Anthropic doesn't claim you should — every patch requires human review and approval before implementation, unchanged from the July launch. Security engineer Kien Pham (@kienbuilds) put the skeptical case bluntly on X: "Can't wait for Mythos 5 to engineer a backdoor in my codebase and then gaslight me into thinking it found the 'smoking gun.'" That's a real risk category worth taking seriously, not just online noise. Any LLM-based scanner — Claude Security included — can produce a confident-sounding false positive, misattribute a root cause, or propose a patch that closes the reported symptom without fixing the underlying flaw. The adversarial verification pass Anthropic added at the July launch is meant to cut false-positive volume, not eliminate it, and a human reviewer who rubber-stamps AI-suggested diffs without independently reasoning through them defeats the entire point of the review gate.
"Does the Mythos 5 upgrade cost extra?" No — Anthropic states scans are billed as standard token usage under a customer's existing Enterprise plan, no separate add-on fee for the Mythos 5 upgrade specifically.
What this means for security teams
If you're already on the Claude Security beta: the upgrade should apply automatically under your existing Enterprise access — no new contract needed for the model swap itself, though Mythos-class access anywhere outside this scan still isn't available to you even as a customer.
If you're evaluating whether to pilot Claude Security for the first time: the "findings only, no direct Mythos access" design is worth understanding before you pitch it internally — it means you're not buying a general-purpose Mythos 5 seat, you're buying a scan feature that happens to use a stronger model underneath. Compare it against your existing SAST/DAST tooling the same way explainx.ai's July coverage suggested: pilot on a subset of your codebase, and watch the false-positive rate specifically, since that's the metric most exposed to the "confident but wrong" failure mode Pham's comment describes.
If your team maintains widely used open-source infrastructure: the Defender Advantage Fund is worth watching even without an application process publicized yet — Anthropic says pilot recipients will be named in the coming weeks, which is the point to check back for actual eligibility details.
If you're a security product vendor: the partner-integration track (Mythos running behind your own product's interface, returning only a defined artifact) is a real distribution path if you don't want to build and maintain your own vulnerability-reasoning model — Anthropic is taking registrations of interest now.
Update — August 22, 2026: This post is a follow-up to explainx.ai's original coverage of the Claude Security public beta launch — read that post for the full three-stage scan/validate/patch pipeline and Claude Code plugin details, which are unchanged here.
Related on explainx.ai
- Claude Security: from scan to fix, public beta launch (July 2026)
- Claude Mythos Preview and Project Glasswing — the original cybersecurity disclosure
- ExploitBench: measuring how far AI can exploit real code
- Why US government banned Fable 5 and Mythos 5 — export control story
- VulnCheck: AI-found bugs aren't exploited at a higher rate than baseline
- Cisco Antares — open-weight vulnerability localization
- Why agent skills are a security risk — and how explainx.ai verifies every skill
- Claude Code commands — complete reference guide
Primary source: Anthropic, "Bringing the cybersecurity capabilities of Claude Mythos 5 to more defenders" (August 21, 2026).
Product and program details reflect Anthropic's August 21, 2026 announcement. Beta features, fund eligibility, and Cyber Verification Program timelines change — verify current status directly with Anthropic before planning a rollout or fund application.