What Rubrik announced
In its October 9 announcement, Rubrik introduced Code Guardian as an expansion of Project Hourglass, its program focused on AI-agent security and recovery. Rubrik says the new tool uses Claude Mythos 5 in a security harness to examine software for multi-step weaknesses. The company describes a private preview for select design partners, not a generally available product.
The distinction matters. The announcement outlines what Rubrik intends the tool to do; it does not provide independently verified detection rates, false-positive rates or customer deployment results. Rubrik itself cautions that availability, features and performance are not guaranteed.
Why test a cloned repository?
Rubrik says Code Guardian will work on a cloned copy of customer code in an isolated, air-gapped environment rather than test a live production system. That design could lower the chance that a red-team exercise disrupts users or exposes a running service. It also creates questions that security teams should resolve: whether the clone includes the dependencies and permissions that shape a real exploit, how sensitive source code is handled, and how findings are reproduced safely.
The stated goal is to follow attack chains across files, services, identity roles and cloud boundaries, verify which paths are exploitable, and rank findings by business impact. If it works reliably, that could help developers focus on more consequential flaws. But a claimed exploit chain still needs evidence and human review before a team changes production code.
Project Hourglass and the partner rollout
Rubrik says it will pair Code Guardian with Rubrik Agent Cloud as Project Hourglass expands. AHEAD, Trace3 and World Wide Technology are joining the alliance, alongside existing partners including Cognizant, Deloitte, HCLTech, LTM, NTT DATA and Wipro. Partners are expected to receive certification and engineering support as deployments develop. Partner participation is not the same as proof of broad adoption or security effectiveness.
What security leaders should ask next
AI coding tools can speed up changes to a codebase. That makes timely security review valuable, but speed alone is not the measure. Teams evaluating AI-assisted red-teaming should ask for reproducible exploit evidence, known limits, false-positive data, clear handling of private repositories and a process for validating fixes. They should also ask how well an isolated test reflects the cloud and identity settings of the deployed system.
Our takeaway: Code Guardian is an interesting proposed way to investigate complex software risks away from production. Its most important proof points are still ahead. The podcast episode and Short above separate the announced design from demonstrated results and explore what credible validation would look like.
Source and listen
Source: Rubrik Newsroom, October 9, 2026. Watch the YouTube Short or listen to the full Daily AI Chat podcast episode.
