ainewsblitz.com

Breaking

Anthropic Brings Claude Mythos 5 to Claude Security for Enterprise Vulnerability Scans

  • Security
  • Foundation Models
  • Sales & Business

Anthropic has moved Claude Mythos 5 into Claude Security, giving Claude Enterprise customers public-beta access to the model’s vulnerability-scanning capabilities without exposing a direct Mythos prompt surface. The service returns structured findings and suggested patches, while a human must review and approve every change. The rollout began on August 21, 2026. An organization owner can enable Claude Security from the admin console, after which users select a GitHub repository at claude.ai/security. The scanner analyzes the codebase and returns each finding with a Common Weakness Enumeration (CWE) category, confidence and severity ratings, and a suggested fix. Scans are billed as standard token usage under an existing Claude Enterprise plan, with no separate add-on. Users can open Claude Code on the web to work on a patch, but that interactive step uses the models already available to their organization; the Mythos scan does not extend access to other Claude surfaces. Every patch requires human approval.\n\nThe product is deliberately designed as a constrained interface rather than a general-purpose endpoint. Claude Mythos 5 runs in the background and returns defensive artifacts such as vulnerability reports and patch suggestions, not raw model output. Anthropic’s stated rationale is that a fixed workflow gives defenders access to difficult security analysis while reducing the opportunity to steer the model toward exploit development. The Claude Security guide says the system scans code in parallel, traces data flows across files, identifies complex multi-component patterns, and puts findings through multi-stage verification before surfacing them.\n\nThere are material boundaries. The hosted service currently supports GitHub repositories only and is intended for code owned by the customer or its organization. Anthropic says scans are stochastic by design, so separate runs can produce different findings; severity configuration is not currently available. The documentation also notes that Claude Security does not offer Zero Data Retention, while the Claude Mythos 5 overview says use of the model requires accepting a 30-day retention policy for safety monitoring.\n\nThe launch extends Anthropic’s cautious rollout of its most cyber-capable model. Claude Mythos 5 is the same underlying model as Claude Fable 5, but with safeguards lifted in selected areas. Fable 5 is broadly available with classifiers that can route cybersecurity, biology, chemistry, and distillation-related requests to Claude Opus 4.8, while Mythos 5 has been restricted to vetted cybersecurity partners through Project Glasswing. Anthropic is now working to place Mythos 5 inside security products used by defenders of hospitals, utilities, financial systems, and the software supply chain, keeping end users inside purpose-built workflows rather than giving them direct model access. This approach is described in Anthropic’s model announcement and its cybersecurity rollout.\n\nEarlier Project Glasswing results explain why the model is significant, but they should not be read as a Claude Security accuracy guarantee. Anthropic reported that approximately 50 partners found more than 10,000 high- or critical-severity vulnerabilities. In more than 1,000 open-source projects, Mythos Preview estimated 6,202 high- or critical-severity vulnerabilities; of 1,752 independently assessed findings, 90.6% were judged true positives and 62.4% were confirmed as high or critical. On CyberGym, Anthropic reported that Mythos Preview reproduced vulnerabilities in 83% of more than 1,500 tasks, compared with 67% for Claude Opus 4.6. Those figures come from the predecessor model and from Anthropic’s own evaluations, not from a published head-to-head measurement of the new Claude Security product. The Glasswing update and Anthropic’s benchmark summary provide the underlying context.\n\nClaude Security also enters a market where GitHub Code Security combines CodeQL with AI-powered detections and Copilot Autofix inside pull requests. GitHub’s AI detections are designed to broaden coverage across languages and frameworks, while its suggested fixes remain part of the review workflow. GitHub’s security update shows the competing emphasis on merge-point integration; Claude Security’s differentiator is repository-level, cross-file reasoning with a separate verification pass.\n\nEarly industry reaction has focused on that trade-off. Some developers viewed outcome-only access as a sensible way to put a powerful security model in defenders’ hands without opening a free-form offensive endpoint. Others criticized the Enterprise-only gate and usage-based access as too narrow for independent security work. The industry reaction reflects the unresolved tension between responsible access controls and broad availability. Anthropic is also launching the $35 million Defender Advantage Fund in Claude credits for organizations securing open-source software, expanding its Cyber Verification Program, and pursuing integrations with cybersecurity vendors. The immediate test will be whether Mythos-powered findings remain valuable after human verification and patch review while customers weigh stochastic results, GitHub-only coverage, retention requirements, and token costs.