Anthropic stated on August 21, 2026 that Claude Mythos 5, the cyber-capable mannequin it has restricted to vetted defenders since April 2026, is now operating vulnerability scans in Claude Safety for Enterprise clients. The identical announcement lays out plans to carry Mythos 5 into safety companions’ merchandise, commits $35 million in credit to a brand new fund for open-source protection, and previews an growth of the corporate’s Cyber Verification Program towards Mythos-class entry.
The strikes mark the third stage of a rollout Anthropic has intentionally saved sluggish. The corporate’s reasoning, specified by the announcement, is that the hazard concentrates the place a person has direct entry to the mannequin and might attempt to steer it towards offensive work. When customers can solely obtain “particular outputs, similar to a patch for a vulnerability or a safety alert,” Anthropic argues, that threat drops significantly, so the growth widens entry to defensive outcomes whereas maintaining guardrails across the mannequin itself.
4 concrete adjustments make up the announcement. Enterprise clients can now run Claude Safety scans on Mythos 5; Anthropic is working with cybersecurity expertise and companies companions to construct the mannequin into instruments defenders already run; the brand new Defender Benefit Fund (0xDAF) will distribute $35 million in Claude credit to organizations securing open-source software program; and the Cyber Verification Program, which presently offers vetted organizations decreased safeguards on Claude Opus and Sonnet fashions, will increase within the coming weeks to cowl broader dual-use capabilities on these fashions, with Mythos-class entry to observe.
How Claude Safety Delivers Mythos 5 With out Opening the Mannequin
Claude Safety is in public beta for Claude Enterprise clients, and Anthropic says scans with Mythos 5 are billed as commonplace token utilization underneath present plans, with no separate add-on. Admins allow the characteristic within the admin console; from claude.ai/safety, a person picks a repository, and the mannequin scans the codebase for vulnerabilities. Every discovering comes again with a CWE (Widespread Weak spot Enumeration) class, confidence and severity scores, and a advised repair. The person then opens Claude Code on the internet to implement the patch, and each patch should be reviewed and authorized by a human earlier than it lands.
The scoping is the purpose. The Mythos 5 scan returns detailed findings moderately than uncooked mannequin outputs, and operating one doesn’t prolong Mythos entry to every other floor: interactive patching makes use of whichever fashions the group already has in Claude Code. Companion integrations observe the identical sample. An finish person of a accomplice product by no means prompts Mythos 5 instantly; a purpose-built interface runs the mannequin within the background for an outlined activity and returns solely the supposed artifact, similar to an inventory of advised patches. Anthropic says it and its companions run abuse-prevention measures to maintain the mannequin inside that scope, and distributors constructing safety merchandise can register curiosity within the integration program.
The accomplice channel builds on an ecosystem Anthropic has been assembling since Might 2026, when it reported that corporations together with Wiz, Palo Alto Networks (PANW ), CrowdStrike (CRWD ), and Accenture had constructed cyber merchandise on Claude Opus. The early figures in that submit are company-reported moderately than independently measured, however they describe the form of the market: Wiz stated its Opus-powered offensive-testing agent runs repeatedly throughout greater than 150,000 manufacturing property per week, and Accenture stated it took safety testing protection from roughly 10% to over 80% throughout 1,600 purposes whereas slicing scan turnaround from 3–5 days to underneath an hour. CrowdStrike’s consulting chief Mark Manglicmot, whose agency sells an Opus-based readiness service, put the accomplice case this fashion: “Frontier fashions like Anthropic’s Claude Opus are giving defenders a functionality benefit that didn’t exist a 12 months in the past, pushing vulnerability administration all the way in which to the left.”
Palo Alto Networks lately enlisted Anthropic, OpenAI, and industrial-control distributors for critical-infrastructure protection, and distributors similar to Harness have begun delivery brokers that scan, triage, and patch vulnerabilities at machine pace.
From Mission Glasswing to a Wider Defender Rollout
Anthropic’s warning with this mannequin class dates to Mission Glasswing, launched on April 7, 2026 with a coalition whose members included Amazon Net Companies, Microsoft, Google, CrowdStrike, and the Linux Basis. The undertaking put Claude Mythos Preview into the arms of a small group securing essential software program, backed by as much as $100 million in utilization credit and $4 million in direct donations to open-source safety organizations. Anthropic stated Mythos Preview had already discovered hundreds of high-severity vulnerabilities, together with some in each main working system and net browser: a 27-year-old remotely triggerable flaw in OpenBSD, a 16-year-old bug in an FFmpeg line that automated testing had hit 5 million occasions with out catching, and a chained Linux kernel privilege escalation. All three had been reported to maintainers and patched earlier than disclosure. This system saved including members over the summer time, with safety corporations like Horizon3.ai becoming a member of the vulnerability hunt in July 2026.
On June 9, 2026, Anthropic break up the mannequin in two for its Fable 5 and Mythos 5 launch. Claude Fable 5 is identical underlying mannequin with security classifiers that route cyber, biology, and distillation-related queries to Claude Opus 4.8 as a substitute; it went to common availability. Claude Mythos 5, with cyber safeguards lifted for vetted customers, went to Glasswing companions at $10 per million enter tokens and $50 per million output tokens, lower than half the value of Mythos Preview. Anthropic stated an exterior bug bounty produced no common jailbreaks in over 1,000 hours of testing, and it has saved tuning the classifiers since: earlier in August 2026 it retuned Fable 5’s biology safeguards, slicing blocked queries 85%. Visitors on Mythos-class fashions carries a 30-day retention requirement, which Anthropic says is used for security functions and never coaching.
Anthropic’s Cyber Protection Push by the Numbers
- $35 million in Claude credit dedicated to the Defender Benefit Fund (0xDAF), for patching reside vulnerabilities in extensively used open-source tasks, automating scanning and patching in methods different tasks can replicate, and approaches that make tasks immune to complete courses of assault
- $100 million in utilization credit dedicated to Mission Glasswing in April 2026, plus $4 million in direct donations: $2.5 million to Alpha-Omega and OpenSSF by way of the Linux Basis, and $1.5 million to the Apache Software program Basis
- $10 per million enter tokens and $50 per million output tokens for Mythos 5, versus $25 and $125 for Mythos Preview
- Greater than 40 further organizations sustaining essential software program infrastructure given entry past Glasswing’s 12 launch companions
- 83.1% versus 66.6%: Mythos Preview in opposition to Opus 4.6 on CyberGym vulnerability replica, as measured and reported by Anthropic itself
The fund extends work Anthropic says Glasswing already began, together with help for coordinated vulnerability-fixing efforts like Akrites and the White Home’s Gold Eagle initiative. It begins with a small variety of bigger pilot grants whereas Anthropic learns what scales.
What Comes Subsequent for Mythos-Class Entry
The near-term milestones are all scheduled. Anthropic says that over the approaching weeks the Cyber Verification Program will increase: defensive capabilities like vulnerability triaging and validation transfer to Mythos-class fashions, and enrolled defenders will see decreased blocks on Opus and Sonnet. Organizations already accepted want take no motion; Anthropic says it should attain out with updates. Preliminary Defender Benefit Fund recipients shall be named within the coming weeks because the pilot grants land, and accomplice integrations carrying Mythos 5 into business safety merchandise are described as early work the corporate expects to increase over time. In parallel, Anthropic says it’s widening Mythos entry by way of Mission Glasswing with U.S. authorities companions, aimed toward operators of essential infrastructure that meet strict safety management necessities.
Claude Safety stays in public beta for Enterprise clients, with Mythos 5 scans reside as of August 21, 2026 and billed as commonplace token utilization underneath present plans.
