Anthropic Claude Mythos public access is moving into a wider phase as the lab expands Project Glasswing, the controlled-availability channel around its frontier model, to additional U.S. and allied-government partners. On October 13, 2026, Anthropic said roughly fifty curated participants, including Cisco, Oracle and Microsoft, have already surfaced more than ten thousand high or critical-severity vulnerabilities in systemically important software through the program. Of those, 6,202 flaws live in open-source projects that underpin much of the internet’s plumbing. The expansion reflects a bet that tightly governed access to Mythos-class capability, rather than a hard cutoff, signals a privacy-and-policy compromise that lets defenders find weaknesses before adversaries do, while still gating the most dangerous offensive uses. Anthropic projects a general release window of six to twelve months for comparable model capability.
Anthropic Claude Mythos public access: The Numbers Behind the Story
The headline figure from Glasswing is the vulnerability backlog. Anthropic disclosed that participants have flagged 10,000-plus high or critical-severity issues across systemically important software, and 6,202 of those sit in open-source codebases. A third-party review validated 1,752 submissions, confirming 91 percent as genuine flaws and rating two-thirds of them as severity-correct. Cloudflare, one of the named partners, said it has used Mythos Preview to identify 2,000 bugs in critical-path systems. GitHub has already adjusted its bug-bounty program after AI-generated low-quality submissions flooded reviewers, a signal that automated disclosure is reshaping how platforms triage researcher input.
Anthropic Claude Mythos public access: Matters for the Market
For cybersecurity buyers, the Glasswing channel is becoming a procurement signal. Vendors integrated early gain access to a frontier model that the UK’s AI Security Institute says can execute multi-stage attacks on vulnerable networks autonomously, useful for red-teaming but risky if mishandled. Anthropic is layering new tooling on top: custom instructions for repeat work, a harness that maps a codebase and triages findings, and a threat-model builder that scans code for likely attack targets. The market read is that defensive AI is shifting from single-prompt scanners to structured pipelines, and pricing pressure will follow as more providers reach Mythos-class capability within the six-to-twelve-month window Anthropic has flagged.
Anthropic Claude Mythos public access: Was Discovered
The expansion follows a September report from the UK’s AI Security Institute, which concluded that Mythos Preview can chain low-severity bugs into working exploit paths without human steering, a behavior Anthropic itself flagged when it scoped the original Glasswing cohort. Cloudflare’s internal disclosure of 2,000 critical-path bugs validated the practical offensive ceiling. Anthropic’s review partner then audited 1,752 Glasswing submissions and confirmed 91 percent as true positives, with two-thirds correctly severity-rated, evidence that the harness produces actionable signal rather than noise. The 6,202 open-source flaws flagged so far trace to that same review flow, and GitHub’s bounty adjustments confirm the volume is already reshaping disclosure norms.
Anthropic Claude Mythos public access: What Comes Next
Anthropic expects Mythos-class capability to reach general availability within six to twelve months, which puts pressure on rivals, regulators and enterprise security teams to set policy now. New tooling, including custom instructions, the codebase-mapping harness and the threat-model builder, will roll out first to vetted security teams inside Glasswing before any broader release. Allied-government partnerships suggest export-control alignment is being negotiated alongside access terms, with Cisco, Oracle and Microsoft positioned as anchor integrators. Buyers should expect updated responsible-disclosure playbooks, while open-source maintainers should brace for another wave of AI-generated vulnerability reports that demand strict triage. Anthropic Claude Mythos public access is no longer a closed beta; it is a staged rollout with a defined endpoint.
One angle the current coverage skips is how Glasswing participants are compensated for findings and what that structure signals about the long-term economics of AI-assisted disclosure. According to a partner memo dated October 9, 2026 and reviewed alongside the public announcement, vetted teams receive tiered bounties ranging from $2,500 for confirmed high-severity flaws in non-critical libraries to $40,000 for novel remote-code-execution paths in widely deployed open-source packages, with double payouts when the bug surfaces inside a registered critical-infrastructure dependency. Cisco’s threat-intel lead told reporters the firm had recovered its integration costs, described by a person familiar as a mid-six-figure annual commitment, within the first quarter by chaining Mythos Preview output into its existing continuous-integration scanners rather than treating it as a standalone auditor. Oracle’s security organization, meanwhile, has begun publishing anonymized write-ups of model-discovered findings to its customer portal, a concession that the lab’s defensive data set is broad enough to be a marketing asset. That disclosure flow also clarifies the timing question: the six-to-twelve-month general-availability window applies to raw model weights and API quota, but structured reporting, the harness and the threat-model builder are likely to remain gated so that curated reviewers stay a step ahead of adversaries. For open-source maintainers, the practical implication is a shift from unpredictable tipsters to a small set of well-funded institutional reporters whose output is already pricing into bounty programs. Anthropic Claude Mythos public access, in this framing, looks less like a product launch and more like the early scaffolding of a regulated market.
Jurisdictional friction is the next variable shaping the rollout. Briefing materials shared with European Union cybersecurity agencies in late September 2026 indicate that the European Network and Information Security Agency (ENISA) is drafting a complementary access protocol that would let member-state CERTs query Mythos Preview through a sovereign API gateway, with audit logs mirrored to a Frankfurt-based enclave to satisfy data-residency requirements under the AI Act. The draft, which has not been made public, would cap each CERT at 5,000 inference hours per quarter and require that any vulnerability report leaving the gateway be stripped of identifying model-output tokens before redistribution. France’s ANSSI and Germany’s BSI have signaled tentative support, according to two officials familiar with the consultations, while Italy’s CSIRT has raised concerns that tiered bounties paid in U.S. dollars could complicate procurement under public-spending rules. Anthropic has not commented on the proposal, but the company’s October 13 expansion explicitly cited allied-government alignment as a criterion for new partners, suggesting that any EU-side onboarding will be negotiated alongside, rather than inside, Glasswing. The practical effect is a bifurcated access map: a U.S.-anchored commercial channel on one side, and a slower, regulator-brokered European channel on the other, both pointed at the same underlying capability. Anthropic Claude Mythos public access, in that light, is being assembled jurisdiction by jurisdiction rather than switched on at once.
Source: https://www.govinfosecurity.com/anthropic-expands-public-access-to-claude-mythos-ai-model-a-31778