Anthropic integrates Mythos 5 into Claude Security to harden enterprise codebases
The AI lab is deploying its most capable frontier model via a mediated interface to give security teams a 'defender's advantage' against cyber threats.
Anthropic has integrated its most capable frontier model, Claude Mythos 5, into the Claude Security vulnerability scanner to help organizations identify and patch software flaws. The move aims to provide security teams with high-end discovery capabilities before similar tools are widely available to malicious actors.
Currently in public beta for Claude Enterprise customers, the Claude Security tool allows organizations to scan their own codebases for vulnerabilities. Rather than providing raw access to the model, the system returns structured findings including Common Weakness Enumeration (CWE) categories, confidence ratings, severity ratings, and suggested fixes. To further support the ecosystem, Anthropic launched the Defender Advantage Fund (0xDAF), providing $35 million in Claude credits to organizations working to patch open-source software.
The push for defensive AI
This deployment is an expansion of "Project Glasswing," an initiative launched in April 2026 that provided a select group of critical infrastructure organizations with early access to Mythos-class models. The strategy reflects Anthropic's attempt to manage "dual-use" capabilities—tools that are equally effective at discovering vulnerabilities for defense and exploiting them for offense.
The potency of the model is evidenced by recent benchmarks. Mythos 5 scored 78.0% on the ExploitBench test for converting discovered flaws into working attacks, significantly outperforming Z.ai's GLM-5.3, which scored 54.4%. The model's offensive potential was further demonstrated during capture-the-flag evaluations conducted with partner Irregular, where Claude Opus 4.7 and Mythos 5 successfully gained unauthorized access to three organizations.
Strategic safety and industry impact
By utilizing a mediated interface rather than a raw chat prompt, Anthropic is implementing a strategic safety layer. This design prevents users from steering the model toward offensive cyber-attacks while still delivering the utility of its scanning capabilities. This approach is central to the "defender's advantage" race, where the goal is to harden global software infrastructure faster than attackers can find new ways to breach it.
What remains to watch
As Anthropic expands its Cyber Verification Program, the industry will be watching how these mediated tools scale across different enterprise environments. While the integration of Mythos 5 provides a powerful shield for Enterprise customers, the tension between providing utility and preventing misuse of frontier models remains a primary challenge for AI labs deploying cybersecurity tools.