Claude Mythos 5 powers Claude Security vulnerability scanner

Claude Mythos 5 and the generally available Claude Fable 5 are described as the same underlying model, with Fable 5 carrying additional safety measures for dual-use capabilities that Mythos 5 does not ^[1]. That gap is exactly why Claude Security is built the way it is: instead of putting Enterprise customers in front of an open prompt box, the product hands back a structured scan result, findings classified by CWE category, severity, confidence rating, and a suggested fix that a human must review before it ships ^[2]. The same model that finds a vulnerability is architecturally prevented from being steered into writing an exploit, because there is no free-text interface to steer.

That containment-by-product-design choice is also, functionally, a monetization mechanism, and the community reaction has picked up on the tension. Reaction on Reddit's ClaudeAI community, after a long comment thread, converged on the read that Anthropic is not handing over raw Mythos 5 access to anyone, enterprise or not; it is a narrowly scoped application with a fixed workflow, a world away from unrestricted model access. A visible undercurrent of that discussion was skepticism that the earlier 'too dangerous for general release' framing around Mythos itself functioned as a setup for this scan-only, paid access tier. Payward Co-CEO Arjun Sethi frames the same product decision more optimistically, arguing frontier AI reverses the usual attacker/defender asymmetry because a model can now read every line of code the way an attacker would, at machine scale ^[3].

Anthropic's headline capability claim is a jump on CyberGym, a vulnerability-reproduction benchmark, from 66.6% for Claude Opus 4.6 to 83.1% for Mythos 5 ^[4]. Interviews around the launch (including a Cloud Security Podcast conversation with Brinqa CSO Brad Hibbert) framed this less as a single event and more as a 'climate change' shift: continuous, compounding AI-driven discovery rather than one-off incidents like Heartbleed or Log4j, with discovery-to-exploit timelines compressing from months to weeks or days, and three chained medium-severity privilege-escalation bugs now sufficient to reach what previously required one critical CVE.

Security practitioners quoted around the launch pushed back on how much that capability actually changes for defenders day to day. Rich Mogull of the Cloud Security Alliance argues the cybersecurity story for the average practitioner has not changed and counsels methodical preparation over alarm ^[6]. Cynthia Kaiser of Halcyon goes further, noting adversaries are already highly effective without AI assistance, breaching some victims in under an hour, which undercuts the idea that AI scanning is the primary threat driver ^[5]. Analysts have also questioned Anthropic's own severity claims: reports of thousands of severe vulnerabilities trace back to expert contractors agreeing with Claude's severity rating in roughly 90% of just 198 manually reviewed cases, and at least one FFmpeg finding that Anthropic's own review concluded was not critical severity ^[5]. The practical effect, according to this critique, is that finding flaws faster does not shrink the real constraint, which is triage and patching capacity, not discovery.

Claude Security's public beta sits downstream of Project Glasswing, the invitation-only program Anthropic launched on April 7, 2026 with eleven founding partners including AWS, Apple, Cisco, CrowdStrike, Google, JPMorganChase, Microsoft, and NVIDIA, giving them access to a safety-loosened Mythos build for security research ^[7]. The program has since expanded substantially, adding about 150 organizations across 15+ countries and, by its own accounting, surfacing more than 10,000 high- or critical-severity flaws across participants; access remains invitation-only, NDA-bound, and approved in consultation with the US government ^[8]. Access pricing has also fallen: direct Glasswing access to Mythos 5 now costs $10 per million input tokens and $50 per million output tokens, down from $25 and $125 respectively for the earlier Mythos Preview ^[4].

The expansion has also pulled in a new sector. Payward, Kraken's parent company, joined Glasswing in mid-August 2026, becoming the first reported crypto firm with access, after more than 40 Bitcoin and crypto companies signed an open letter asking for frontier-model access specifically for defense ^[3]. That sequence, an open letter followed by Payward's admission into the program, is a concrete instance of sector-wide lobbying preceding entry into Glasswing.

Coverage of the private-preview period before this launch reported concrete downstream effects of faster vulnerability discovery: Mozilla fixed 271 Firefox vulnerabilities after testing Mythos, and researchers used it to chain together vulnerabilities that bypassed macOS protections. Those results are part of why observers describe the shift as a change in exploit timelines rather than a one-time capability bump, three chained medium-severity bugs now doing the work of a single critical CVE, and remediation windows that used to run 30, 60, or 90 days looking increasingly out of step with how fast attackers can move once a class of bug is public.

That compression has already reached regulators, at least in India: reporting describes India's CERT-In directing organizations to patch within 12 hours, the Reserve Bank of India holding consultations, and Punjab National Bank earmarking additional cybersecurity budget, all attributed to the faster exploit timelines associated with AI-assisted vulnerability discovery. The caveat attached to that reporting is important: hardening driven by tools like Mythos applies going forward, not retroactively, so legacy and already-deployed systems remain exposed even as newly scanned codebases get patched faster. Journalist Siddharth Jindal's framing of the underlying tension is worth keeping in view: Mythos finds vulnerabilities rather than creating them, but the same capability that identifies a flaw implies the capability to help patch it as well, which is the defensive promise and the dual-use risk in the same sentence.