# Claude Mythos Preview and AI-assisted cybersecurity Anthropic frames Mythos Preview as a leap in security work and created Project Glasswing to arm defenders before attackers catch up. Last week Anthropic announced something every technical leader should read: the **Claude Mythos Preview**, a general-purpose model that shows unprecedented capability at **finding and exploiting software vulnerabilities**. Alongside it came **Project Glasswing**, an initiative to put those capabilities in the hands of defenders before malicious actors build equivalent tooling. The move is calculated. The implications run deep. ## What Mythos Preview actually does The difference is not "finding bugs". SAST and DAST tools have done that for years. What changes is the kind of vulnerability the model can identify, and how deep the analysis goes. Traditional tools like **Snyk**, **SonarQube** and **Semgrep** work from static rules and known patterns. They are excellent at catching obvious SQL injection, dependencies with a published CVE, or unsafe use of cryptographic functions. They fail on business logic vulnerabilities, subtle race conditions, authorization flaws that depend on the application's full flow, and exploit chains that combine several individually harmless weak points. Mythos Preview operates at another level. According to the benchmarks Anthropic published, the model can: - **Identify authorization logic vulnerabilities** that require understanding the full application context, not an isolated file - **Build exploit chains** by combining smaller flaws (a race condition here, an IDOR there, a validation bypass somewhere else) into working attacks - **Analyze compiled and obfuscated code**, including reverse engineering binaries to find entry points - **Generate working proofs of concept**: instead of flagging that something might be vulnerable, it shows how to exploit it That puts the model closer to a senior security researcher than to a linter on steroids. ## Project Glasswing: containment before release Anthropic understood the obvious: a model that finds zero-days this efficiently cuts both ways. **Project Glasswing** is the answer. The initiative gives early access to defense teams in critical infrastructure, maintainers of widely used open source projects, and national CERTs. The company does not plan to make Mythos Preview broadly available in the near term. The restriction is deliberate. The same model that helps audit a repository can, in the wrong hands, automate exploit discovery at industrial scale. The dual-use risk is not theoretical. It follows directly from training models that reason deeply about code. ## What this means for technology leaders Three points that demand action now: 1. **Shorten the patch cycle.** The window between CVE disclosure and active exploitation has already shrunk from weeks to days. With models that can generate exploits automatically, it will shrink to hours. Quarterly update processes are a death sentence. If your organization still depends on manual patch cycles, the accumulated risk is already unacceptable. 2. **Automated triage with context.** The volume of findings AI tools will produce is orders of magnitude beyond what any team can process by hand. Without smart prioritization, based on real exposure, asset criticality and confirmed exploitability, the result is paralysis by noise. 3. **Vendors as an extension of risk strategy.** AI labs and cloud providers are no longer just infrastructure suppliers. Which model to use, with which guardrails, under which responsible disclosure policy: all of that is now a risk management decision. Treat it as one. For companies in regulated sectors, the debate is not "whether to use AI in security". It is **how to integrate models with an audit trail, environment segregation and responsible disclosure policies**. That architecture is exactly what we build in [cloud transformation](/capabilities#cloud) and [software engineering](/capabilities#engineering) projects. ## What changes for SOC teams and vulnerability management programs If you lead a SOC or own a vulnerability management program, prepare for a structural shift. **First: the volume of reported vulnerabilities will explode.** Models like Mythos Preview will feed scanning pipelines that surface classes of flaws current tools simply ignore. Your backlog will grow. The temptation will be to ignore low-severity findings. Resist it: exploit chains start exactly at the points nobody prioritizes. **Second: analyst skills have to change.** SOC analysts who only know how to run predefined playbooks will lose relevance fast. The skill that matters now is reading the output of AI models, validating whether an exploit chain is viable in your specific infrastructure, and translating technical findings into business decisions for the board. **Third: traditional metrics go obsolete.** MTTR (Mean Time to Remediate) measured over public CVEs no longer captures real risk. You need to measure semantic analysis coverage, the false-negative rate of your current tools versus AI-assisted analysis, and the time between internal discovery and effective fix. Vulnerability management programs that do not absorb AI-assisted analysis in the next 12 to 18 months will run blind to entire classes of vulnerability. ## Where Uranus stands Our work in [GRC](/capabilities#grc) and [AI agents](/capabilities#agents) already starts from the premise that language models are both an attack surface and a defense surface. We help organizations build the governance, architecture, and processes to use these capabilities without opening new risk vectors. --- **Read also:** [Hack on China's supercomputer exposes 10 PB of military data](/blog/nscc-supercomputer-hack-cybersecurity-lessons) Recommended reading in full on Anthropic's blog: [Assessing Claude Mythos Preview's cybersecurity capabilities](https://red.anthropic.com/2026/mythos-preview/).