Security professionals live in a time of new AI realities. Threat discovery, once the main focus of security assessment, is now fast and abundant. Breaches that used to evolve over weeks now happen in seconds. Vulnerability chains, long invisible to the best security teams and developers, are instantly surfaced by powerful new frontier models. Across a large codebase, the model has the potential to find large volumes of potential security issues.
In June, Rubrik joined Project Glasswing, which gave our engineers access to Anthropic's Mythos Preview. After a month of working with the model that focused the industry on new challenges, I'm confident that we can find effective ways to meet the new challenges of the AI era – and to improve cyber resilience.
In particular, we brought together a multifunctional 'tiger team' from engineering and infosec, prioritizing high-fidelity threat discovery and elimination within our own systems, and automating processes wherever possible. The focus was to build an effective harness (the software layer wrapping the AI model) that manages tool calls and checkpoints, adds business context, security context, and trust boundaries.
If attacks now happen at AI speed, solutions will also have to be produced at that speed.
In practice, that means the work doesn't end with discovery. At these volumes, every stage downstream becomes a constraint — and the hardest engineering problems turned out not to be the ones we started with.
A High-Fidelity Approach
From Rubrik’s work using Mythos as part of Project Glasswing, our first instinct was to treat new discoveries as a capacity problem: bigger queue, more reviewers. But we realized that adding capacity was not the fix; we had to rethink what the pipeline was doing.
What actually worked was treating this as an architecture problem. Instead of asking "how do we review our findings?" we considered "how do we build a system where only the right findings reach an engineer?”
That shift changed how we thought about the harness. Business context, security context, the specific threat model of a company where many customers rely on us as a last line of resilient recovery: none of this could live in a prompt. Prompts drift. Harness architecture doesn't. When context is structural, it applies consistently across every scan and every pass, without depending on anyone getting the wording right each time.
Whole-repository scans came first, no assumptions, every file. We progressively targeted passes, seeded by what early rounds revealed about where the real patterns were. The result was a significant reduction from findings to validated, priority issues. High fidelity is what makes the downstream remediation viable.
Automation and Humans in the Loop
What has surprised my team most about building a remediation pipeline wasn’t the automation. It was how much of the engineering work we’ve done is going into deciding what not to automate.
The underlying problem is structural. Human-driven remediation will not keep pace with AI-speed discovery — that’s not a resourcing problem you can hire your way out of. The only answer is building automation that can match the rate at which potential vulnerabilities are surfaced, but doing it in a way you can actually trust.
The automated path covers a deliberate subset of potential vulnerability classes — the ones where machine remediation is reliable and well-scoped. Everything outside that set gets structured plans and richer context, but human judgment owns the fix. In security, trustworthy automation and maximum automation pull in different directions. We chose the former.
In building a harness for Mythos Preview, it became clear that AI raises the demand for engineering rigor in the systems surrounding it, not lowers it. The taxonomy, the verification architecture, the decisions about where humans stay in the loop: these weren’t incidental to the work. They were the work.
The Vigilance of Many
This work was only possible thanks to the effort of many people. Other companies in Project Glasswing were generous in sharing their early lessons. Anthropic itself has been a great partner, keen to learn and adapt. My team at Rubrik, set up ahead of time to be relatively small and focused, prioritizes well, and is executing brilliantly.
Earlier this year the release of Mythos was a wakeup call to the industry. We have started to answer it. In the long term I believe we will create better software and better security procedures as a result of this effort—that is the reward of persistent vigilance.