Rubrik has introduced two AI-focused security capabilities aimed at software vulnerability analysis and agentic incident response, combining a Claude-based code red-teaming service with Model Context Protocol access to Rubrik’s security data and workflows.
Rubrik Code Guardian uses a custom harness built around Anthropic’s Claude Mythos 5 to analyze an air-gapped copy of a customer’s source-code repository. Separately, Rubrik MCP gives connected AI agents programmable access to supported Rubrik Security Cloud capabilities through the Model Context Protocol.
Both products remain pre-release. Code Guardian is accepting select design partners for private preview, while Rubrik MCP is in private preview for existing customers and is targeted for general availability in October 2026.
Code Guardian focuses on exploit chains rather than isolated alerts
Rubrik said Code Guardian is designed to reason across files, services, authentication patterns and cloud boundaries to find multi-step vulnerability chains. The service validates findings for exploitability before surfacing them and prioritizes issues according to exploitability, potential blast radius and business criticality.
The model operates against a secure copy rather than a live source repository or production system. Confirmed critical findings can be pushed into developer workflows through Jira or GitHub issues with file-level remediation guidance. Rubrik also links the process with recovery workflows so customers can restore a known-good codebase after a security incident or problematic build.
Michael Moore, cybersecurity lead at Anthropic, said the companies are working to put Claude Mythos 5’s cyber capabilities in defenders’ hands so attack paths can be identified and validated before attackers use them.
The announcement does not provide independent benchmark results comparing Code Guardian with conventional application-security scanners, and the service is not yet generally available. Its ability to reduce false positives or identify previously missed chains will therefore need to be assessed as customers test it in production-like environments.
MCP gives agents a route into Rubrik workflows
Rubrik MCP addresses a different part of the agentic-security stack. It exposes the Rubrik Security Cloud API schema to connected agents, allowing them to invoke available data, identity and application capabilities subject to the customer’s permissions.
Teams can save multi-step recovery or compliance workflows and make them reusable across supported AI clients. Rubrik said the MCP implementation maintains role-based access-control parity, configurable permissions and guardrails aligned with the OWASP MCP Top 10.
The company said Rubrik AI is now used by one-third of its global customers. That figure is company-reported and does not indicate how many customers are using the new MCP capability, which is still in private preview.
Rubrik said its agent architecture was developed with Anthropic’s teams to reduce response latency and improve multi-step reasoning. The MCP announcement follows a broader industry effort to let AI agents interact with enterprise systems through standardized interfaces while retaining identity, authorization and audit controls.
Agent access raises governance questions
Giving an AI agent direct access to security and recovery systems can shorten the path from analysis to action, but it also increases the importance of permission boundaries and traceability. A poorly scoped agent could otherwise reach sensitive telemetry, trigger inappropriate workflows or make an automated mistake during an incident.
That issue is closely related to concerns raised in a recent TNGlobal INSIDER contribution from Rubrik, which argued that organizations need better observability and traceability as they deploy AI agents.
For Code Guardian, the practical question is whether frontier-model reasoning can consistently identify exploitable chains without overwhelming developers with findings. For Rubrik MCP, adoption will depend on whether organizations can give agents useful access while preserving the same controls they expect from human operators and existing automation.