Anthropic Deploys Mythos Model to Expose Vulnerabilities

Anthropic says its new Claude Mythos Preview model has discovered thousands of software vulnerabilities, including decades-old flaws that eluded prior testing. Rather than a public release, Anthropic is running a restricted program called Project Glasswing that shares the model with major cloud, security, and infrastructure firms, including Amazon, Microsoft, Apple, CrowdStrike, and others. The company is supplying roughly $100 million in compute and forming a multi-stakeholder coalition to responsibly surface and patch defects. The capability compresses time to discovery, creating defensive advantages but also raising acute risks if adversaries obtain similar tooling or if disclosure outpaces maintainers ability to patch.
What happened
Anthropic says its new, unreleased model, `Claude Mythos Preview`, can identify software vulnerabilities at an unprecedented scale, finding thousands of high-severity flaws across operating systems, browsers, and widely used applications. In response Anthropic launched Project Glasswing, a restricted alliance that shares the model with about 40 corporate, infrastructure, and security partners and provides roughly $100 million in compute credits to accelerate responsible disclosure and remediation.
Technical details
The model is described as a general-purpose next-generation Claude-family system with unusually strong code and reasoning capabilities. Anthropic reports that Claude Mythos Preview:
- •Finds subtle logic and memory-corruption issues that escaped traditional fuzzing, static analysis, and manual review
- •Produces exploit chains and actionable proof-of-concept hints that shorten triage and patch cycles
- •Shows measurable gains on open security benchmarks, including improved performance on CTI-REALM, per Anthropic and partner statements
Participating companies include Amazon, Microsoft, Apple, network and chip vendors such as Cisco and Broadcom, and security firms like CrowdStrike and Palo Alto Networks. Anthropic has circulated a restricted build rather than a public API to limit misuse. Early results include discovery of vulnerabilities reportedly as old as 27 years, and several bugs that remained undetected after millions of prior tests.
Context and significance
The speed and quality with which a single advanced model can find latent vulnerabilities changes the economics of both offense and defense. For defenders, AI-driven discovery can dramatically reduce blind spots, prioritize patches, and scale security testing across large codebases. For attackers, similar capabilities would compress the time from bug discovery to weaponization, increasing the velocity of zero-day exploitation. That dual-use property explains Anthropic's choice to limit distribution and to create an unprecedented cross-industry coalition.
Project Glasswing illustrates a new operational pattern: competitors pooling access to a high-risk capability to protect shared infrastructure. This mirrors other industry responses to systemic risks, where collaboration substitutes for unilateral disclosure. The alliance model also recognizes the ecosystem problem: many critical maintainers and small vendors lack capacity to absorb flood-stage vulnerability disclosures, creating a remediation bottleneck that could be exploited.
Risks and operational implications for practitioners: Claude Mythos Preview can generate human-actionable exploit paths, so defenders must assume increased discovery rates across the software supply chain. Security teams should plan for:
- •Accelerated triage workflows and prioritization rules to handle higher vulnerability volumes
- •Increased investment in patch automation, CI/CD security gates, and dependency hygiene
- •Expanded coordination with vendors, open-source maintainers, and industry-sharing bodies
What to watch
Regulators and government agencies will weigh whether to restrict models used for binary/exploit generation. Watch for expanded public-private programs, new disclosure norms, and vendor policies that control model-based security testing. The immediate challenge is operational: can maintainers absorb a surge of high-quality findings before adversaries replicate or obtain similar models?
Bottom line: This is a step-change in tooling for vulnerability discovery and a wake-up call for security operations. Controlled deployment and multi-stakeholder coordination reduce immediate risks, but the underlying capability will diffuse. Organizations must treat model-driven vulnerability discovery as a structural shift that requires process, tooling, and policy adaptations.
Scoring Rationale
A frontier model that materially changes vulnerability discovery and forces new industry coordination is industry-shaking. The restricted release and major partners limit immediate fallout, but the dual-use risk and operational burdens on maintainers make this highly important for practitioners.
Practice interview problems based on real data
1,500+ SQL & Python problems across 15 industry datasets — the exact type of data you work with.
Try 250 free problemsStep-by-step roadmaps from zero to job-ready — curated courses, salary data, and the exact learning order that gets you hired.

