Rubrik harnesses Anthropic's Claude Mythos 5 to red-team enterprise code

Cybersecurity firm Rubrik has launched Code Guardian, a product that deploys Anthropic's Claude Mythos 5 large language model inside a secure, air-gapped environment to probe enterprise source code for vulnerabilities before attackers can exploit them. The announcement, made in Dubai on 17 September 2026, positions Rubrik at the sharpest edge of what security practitioners are calling agentic cyber resilience: the use of autonomous AI systems not merely to monitor threats, but to actively simulate them.

The product operates on an isolated clone of a customer's code repository, keeping the analysis entirely separate from live production systems. Claude Mythos 5 then reasons across files, services, authentication flows, and cloud boundaries to surface multi-step vulnerability chains, the kind of chained, context-aware exploits that traditional static-analysis scanners routinely miss. Confirmed critical findings are automatically routed into developer workflows via Jira or GitHub, with file-level remediation guidance attached.

The attacker-defender arms race goes agentic

The strategic logic behind Code Guardian reflects a broader inflection point in enterprise security. Frontier models have compressed the time and expertise required to discover and chain software vulnerabilities. What previously required a skilled adversary spending days manually tracing attack paths can now be delegated to an AI agent operating at machine speed. Rubrik's co-founder and CTO, Arvind Nithrakashyap, framed the product explicitly in those terms: "Frontier models are advancing at a fast pace. In the wrong hands, these models can be used to discover, chain, and exploit vulnerabilities at machine speed. To move just as fast, we are using frontier AI to scale vulnerability detection faster than our traditional code scanning tools."

Anthropic's cybersecurity lead, Michael Moore, confirmed the partnership is part of a deliberate effort to place Claude Mythos 5's offensive capabilities in defenders' hands, with the explicit goal of validating attack paths before adversaries reach them. The air-gapped architecture is not incidental: it is the mechanism that allows a frontier model to be pointed at sensitive intellectual property without exposing that code to external infrastructure or the model provider's own systems.

Cross-sector read-across: who else is in the crosshairs

The convergence of frontier AI and offensive security tooling carries implications well beyond any single cybersecurity vendor. For the financial services sector, where regulators in the EU and UK are tightening software resilience requirements under DORA (the Digital Operational Resilience Act), the ability to demonstrate proactive, AI-validated vulnerability testing may soon shift from competitive differentiator to regulatory baseline. Banks and insurers running complex, multi-cloud architectures face precisely the chained-vulnerability risk that Code Guardian targets.

The defence and critical-infrastructure sectors face an equivalent pressure from a different direction. Nation-state actors and advanced persistent threat groups have been early adopters of AI-assisted offensive tooling; sovereign security agencies in the UK, US, and GCC have each flagged this acceleration in recent threat assessments. A commercially available red-teaming service built on the same class of frontier model now available to adversaries is, in effect, a democratisation of the defender's capability.

For the broader technology investment landscape, the Rubrik-Anthropic collaboration illustrates a structural shift in how AI foundation-model providers are monetising their most capable systems. Rather than selling raw API access, Anthropic is co-developing specialised harnesses with domain-specific partners, capturing value at the application layer while maintaining control over the model's most sensitive capabilities. That model has direct parallels in the way cloud hyperscalers packaged security tooling in the early 2010s, and it signals that the frontier-AI market is beginning to stratify between commodity inference and high-trust, domain-validated deployments.

Code Guardian is currently in private preview with select design partners, with no general availability date disclosed.