Rubrik found more bugs than its engineers could fix

Rubrik rebuilt its vulnerability workflow around Mythos Preview after one month of use, having concluded that no amount of additional human reviewers could keep pace with the rate at which the model surfaces issues. CTO and co-founder Arvind Nithrakashyap described the process to The New Stack.

At a glance

• The company's answer was a software harness around the model that manages tool calls and checkpoints and injects business context, security context and trust boundaries before findings reach an engineer.
• Rubrik joined Project Glasswing in June, when Anthropic widened the invitation-only program to roughly 150 organizations across 15 countries, and only vetted partners on that list can use Mythos Preview.
• Automated remediation was deliberately confined to a narrow subset of vulnerability classes where machine fixes are reliable and well defined, and everything else Mythos surfaces is routed to engineering teams.

Why it matters The bottleneck Rubrik describes appears to be structural rather than specific to one company: detection scales with model throughput, while remediation still depends on engineers reading code and deciding what a fix should do. That gap shifts the value toward the surrounding software layer, the harness, the triage passes and the routing rules. For teams weighing similar access, headcount likely reads as the wrong lever.

Rubrik's first instinct was to hire more human reviewers

Mythos began flagging complex vulnerability chains that Rubrik's existing security tooling and review methodology had not caught, including relationships between components spread across large codebases that neither a conventional scan nor a single engineer's review would surface, according to Nithrakashyap.

The volume created a prioritization bottleneck. Nithrakashyap says the team initially read the readout of potential issues as a capacity problem and weighed adding more human reviewers, then dropped the idea once it became clear that human-driven remediation could not match AI-speed discovery.

Instead of scaling the reviewer base, the company assembled a multi-functional engineering and infosec team charged with high-fidelity threat discovery and elimination, leaning on automation wherever it could rather than on additional headcount. The stated goal was a software layer around Mythos that reduces the number of findings reaching engineers for review and remediation.

A whole-repository scan feeds progressively narrower passes

The harness manages tool calls and checkpoints and adds business context, security context and trust boundaries, in Nithrakashyap's description. Adapting, he says, requires integrating structural context directly within the harness and building systems that categorize and filter findings into actionable insights for remediation.

In practice, Rubrik first runs Mythos across an entire repository and uses those initial findings to shape progressively more targeted passes. The later passes strip out noise so that only high-quality findings reach the appropriate teams and are prioritized there. Only after those targeted passes were in place, Nithrakashyap says, could the engineering team build a workflow that relays high-priority, actionable findings and keeps remediation manageable.

Automated remediation is limited to a predefined set of vulnerability classes

Nithrakashyap frames the harder problem as deciding what not to automate, calling out an inherent conflict between trustworthy automation and maximum automation. According to him, the question came up frequently, and Rubrik resolved it by capping the scope of machine-driven fixes rather than maximizing them.

To preserve trust, automated remediation was limited to a deliberate, tightly-scoped subset of vulnerability classes where machine-driven fixes are highly reliable and well defined. Anything Mythos surfaces outside those categories is routed to engineering teams, where human judgment owns the final fix.

His conclusion after the first month is that most existing engineering workflows are not built for the model's speed of discovery, and that the systems surrounding it carry more of the load than before. Anthropic describes Mythos as its most capable model.

Mythos has shown us that AI actually increases the demand for engineering rigor in the systems that surround it.

What's next Whether the automated path widens beyond the current subset of vulnerability classes is an open question, as is how much of the findings volume the harness filters out before it reaches engineers. Project Glasswing remains invitation-only, so access to Mythos Preview is confined to vetted partners, and no date has been given for broader availability of the model.