Rubrik’s lessons from one month with Mythos Preview
Rubrik, a security and AI company, joined Anthropic's Project Glasswing and spent a month experimenting with Mythos Preview. The AI model surfaced more vulnerabilities than Rubrik's engineering team could manually review, exposing a critical bottleneck in human-in-the-loop workflows. To cope, Rubrik built a middleware layer that manages tool calls, checkpoints, and context, automating triage to filter findings before human escalation. This shift from human-driven to automated remediation reflects a broader need: agentic AI accelerates discovery beyond what manual validation can handle. For operators, the takeaway is that successful AI deployment depends on constructing robust orchestration software, not just deploying models. A caveat: the report is based on a single company's early preview experience, so generalized conclusions about agent scalability remain preliminary.
Rubrik’s lessons from one month with Mythos Preview
Rubrik’s lessons from one month with Mythos Preview. Anthropic’s model uncovered more vulnerabilities than the security and AI company’s engineers could handle.
Key takeaway
Automation of triage and context management must precede human review to avoid creating bottlenecks when AI agents operate at machine speed.
What happened
After joining Project Glasswing, Rubrik gained access to Anthropic's Mythos Preview and tested it for a month. According to The New Stack, Anthropic's model uncovered more vulnerabilities than Rubrik's engineers could handle, forcing the company to reconsider its review pipeline.
To manage the volume, Rubrik built a software layer that controls tool calls, checkpoints, and context, filtering findings before they reach analysts. This represents a deliberate pivot from human-driven security review to automated remediation, highlighting the operational ceiling of manual inspection when AI accelerates discovery.
Evidence
Anthropic's Mythos Preview model uncovered more vulnerabilities than Rubrik's engineers could handle.
The New Stack AI · attributed
Anthropic’s model uncovered more vulnerabilities than the security and AI company’s engineers could handle.
Rubrik built a software layer that manages tool calls, checkpoints, and context to filter findings.
The New Stack AI · attributed
they built a software layer that manages tool calls, checkpoints, and context to filter finding…
Rubrik's shift from human-driven to automated remediation highlights the bottleneck of human-in-the-loop scaling for AI agents.
The New Stack AI · attributed
Rubrik's shift from human-driven to automated remediation in Anthropic's Mythos preview highlights the critical bottleneck of human-in-the-loop scaling for AI agents.
Why it matters
The incident demonstrates that as AI agents achieve discovery speeds beyond human capacity, organizations must redesign their operational workflows around automated validation and context management, or risk being overwhelmed by the very outputs meant to help them.
Limits and uncertainties
The report is based on a single company's early preview experiment, not a production deployment.
The specific performance metrics of Mythos Preview are not disclosed in the article.
The described software layer is proprietary to Rubrik and details are limited.
Practical implications
Operators should assume that AI agents will generate findings at a pace exceeding manual review capacity, and plan automation for triage early.
Build middleware that centralizes tool calls and context to preserve trust and auditability when automating remediation.
Consider human checkpoints only for high-confidence, high-impact actions rather than every event.
What to watch
Anthropic publicly documenting Mythos Preview's scalability and integration patterns.
Rubrik publishing more detailed benchmarks or architectural specifics of its middleware.
Other Project Glasswing participants reporting similar operational bottlenecks.
