This short article is for technology executives defending critical services.
Anthropic tells us that it has built an AI model so capable at finding and weaponizing software vulnerabilities that it won’t release it publicly. Access is being rationed through Project Glasswing to a curated group of technology companies and financial institutions.
The numbers are already ugly
Median time to exploit a disclosed vulnerability is now a single day. Average time to remediate a critical one is over sixty. Nearly a third of 2025’s exploited CVEs were weaponized on or before public disclosure. Frontier models widen that gap, not narrow it. And commoditization is coming with them - Android exploit listings on dark-web markets are up more than 1,000% year-on-year. If Mythos-class capability leaks or replicates into open weights, the marginal cost of a working exploit collapses.
However this plays out - whether offensive and defensive AI co-evolve in an arms race or capability simply diffuses — the real work is unglamorous: patch cycles that need to collapse from weeks to hours without breaking production, release trains that push coordinated fixes under duress, and major incident management that coordinates humans, AI agents and automation on the same task list — with evidence every step ran correctly and the true learnings for next time.
Runbook-Ops, not ticket-ops and not chat-ops
This is complex work. It’s Runbook-Ops, not ticket-ops and not chat-ops. ITSM logs the incident. Call-escalation wakes the team. Neither tells them exactly what to do, nor proves it got done. That is the gap a Mythos-era threat model exposes - and it is where the emerging defensive moat actually sits. The timestamped, graph-structured record of how a complex incident actually unfolds - which task, by whom, in what sequence, with what result - does not exist in any log, ticket or chat thread.
Assurance has to come from the response process
The bitter lesson underneath all of this that you already know: you need a strategy that has a lot of capability after the point of prevention fails. It’s astounding that many providers in the space stop at prevention and leave you on your own after that. Assurance has to come from the response process.
For executives in financial services, energy, healthcare and telecoms, the moat is no longer “we tested it.” It is “we can respond to the unknown, faster than the attacker can weaponize it, with a process the regulator can audit.”
That’s the bar now.
Frequently asked questions
What is the defender’s asymmetry problem?
Attackers can now weaponize a vulnerability far faster than most enterprises can remediate it, and frontier AI models widen that gap rather than narrow it.
What is Runbook-Ops?
Coordinating humans, AI agents and automation on the same task list during complex work, with evidence that every step ran correctly. It differs from ticket-ops and chat-ops, which record or discuss the work but do not direct it or prove it was done.
Why is prevention not enough for cyber resilience?
Prevention will sometimes fail. Assurance has to come from a response process that can handle the unknown faster than an attacker can weaponize it, and that a regulator can audit.
See how Runbook-Ops handles a major incident: cutover.com/book-a-demo
Originally published on LinkedIn, April 2026; exploit timing updated October 2026.