Rubrik rebuilt its review pipeline rather than hire triagers for an AI's findings pile

Rubrik has rebuilt its code review pipeline after using Anthropic's Mythos Preview model to find a significant number of potential security issues in its own codebase, according to SC Media's summary of reporting by Silicon Angle ^[1][2]. The model surfaced so many vulnerabilities that the company overhauled the review process instead of scaling up human review, which is the inverse of the usual response to a noisy new scanner ^[3].

The access came through Anthropic's Project Glasswing, an early-access program for Mythos Preview, a model built to find software flaws and construct attack chains ^[4]. Rubrik did not simply point it at the repositories and read the output. It stood up a dedicated team to build a harness around the model, feeding in business and security context to manage how it ran ^[5]. The scanning itself was staged: a broad first pass across entire code repositories, then narrower passes targeted at the patterns the first pass turned up ^[6].

That sequence is the actual product of this exercise. A model that reports everything it can justify is not useful to a team that has to fix things in priority order, and the harness is where the judgment about what matters to this company, in this codebase, gets encoded. Rubrik says the funnel worked, reporting a substantial reduction from raw findings to prioritized issues, but it disclosed no numbers ^[7]. Without them, the claim cannot be checked in either direction: a filter is judged by what it wrongly discards, and nobody publishes their false negatives.

The more durable decision is the one Rubrik made about the far end of the pipeline. Its stated takeaway was that not every part of vulnerability remediation should be automated, and that for critical issues it chose trustworthy automation over maximum automation ^[8]. In practice that means the model finds and ranks, and humans still own the fix where the fix can break production. Teams that automate the discovery side and then let the same tooling write and merge patches to critical paths are making a different bet, and Rubrik's version is the more conservative one.

The context from Anthropic is that Glasswing partners have collectively identified more than 10,000 high- or critical-severity vulnerabilities, and that patching rates have been a problem across technology, power, water, and healthcare ^[9][10]. That aggregate is the only hard number in the account; Rubrik's own counts are not public ^[11]. It also frames the harder question. Rubrik could assign engineers to build a harness and redesign a pipeline. A water utility that receives a list of critical findings has the same discovery capability and none of the same absorptive capacity, and a finding nobody can patch is a liability transfer, not a remediation.

What to watch: whether Rubrik publishes before-and-after numbers on findings, prioritized issues, and time to patch; whether the harness is described in enough detail to be reproduced by anyone else; and whether Glasswing reports patch rates by sector rather than aggregate discovery counts, since that is the metric that shows whether the program closes holes or only inventories them.