The threat we learned about through Project Glasswing is real. Here is what we think CISOs need to do about it.
Anthropic's Mythos AI model found zero-day vulnerabilities across every major operating system, every major web browser, and critical infrastructure software running power grids, hospitals, and airlines worldwide. Not theoretical vulnerabilities. Working exploits. The model chained four browser vulnerabilities together and produced a JIT heap spray that broke out of both the renderer and OS sandbox. It found remote denial-of-service attacks, smartphone firmware bugs, desktop privilege escalation chains. Anthropic assembled a coalition of 40+ technology organizations with $100M in model credits to coordinate a response. The window between vulnerability discovery and weaponization with this tool is hours and days and not weeks and months.
The Cyberstarts 2026 Cybersecurity Trends Report predicted it. Glasswing delivered it ahead of schedule. Every major thesis in the report — machine-speed conflict, AI as the dominant threat accelerant, identity and data as the primary attack surfaces, resilience metrics replacing prevention metrics, AI vs. AI as the only credible defensive posture — shows up directly in what
Glasswing's revelations compressed findings into a single week rather than a three-year arc. The report cited attack timelines collapsing from 285 minutes to 72 minutes as evidence of where the conflict was heading. Glasswing is the destination. The one number that ties it together is the same 40-to-1 ratio at the center of both: 204 days to patch, 5 days to weaponize, and a confidence gap where only 17% of CISOs believed their programs could hold. The report explained why that gap existed. Glasswing just made it impossible to ignore.
Security researchers have been worried about AI-assisted attacks for years and what is worse is that many enterprises are not prepared for what is next.
01 · Context
Why Mythos Is Different
AI has been part of security workflows for a while. Anomaly detection, log correlation, threat intel feeds. Good at pattern matching. Not great at discovery.
Mythos is something else. It works like a researcher, the kind with deep offensive skills and no sleep requirements. Anthropic had to pre-brief over 40 organizations before releasing anything publicly because the model was generating exploits sophisticated enough to matter in the real world.
The underlying math was already ugly before this week. The 2025 DBIR put the average patch time at 204 days. Average time to weaponize a known vulnerability: 5 days. A 40-to-1 ratio that's been sitting in the background of every vulnerability management conversation for years. Mythos doesn't close that gap. It blows it wider.
The question isn't whether you get breached. It's whether you find it fast, contain the damage, and get back to work before it matters.
02 · Shift
This Cuts Both Ways
Before the doom sets in: we know that the same capability that found those vulnerabilities can be leveraged by defenders to help better understand risk. Most serious security organizations are already thinking about running AI tools “offensively” against themselves as a regular operational discipline. We have been talking about it as AI vs. AI. That's the right approach, and now it's urgent. The CISOs and security leaders who understand and can execute effectively in this new operational paradigm are going to be in high demand.
03 · Action
Where the Leverage Is
Here's where we see the most traction for CISOs figuring out what to do.
Adopt AI Powered Assessment Capabilities NOW
The annual pen test is functionally dead as a security control. What matters now is how fast you can run continuous, AI-driven offensive simulation against your own environment. Real-time visibility into actual exploitability beats a quarterly report, every time. A measured understanding in real time on what risks are managed with compensating controls and which are not is critical in this dynamic world.
CVSS scores are mostly noise at this point. Exploitable platforms (adopt EPSS vs CVSS) that are not protected by compensating controls that expose critical systems or data is where I'd focus first.
Know Exactly What a Breach Could Impact
If undiscovered vulnerabilities are now table stakes, the question shifts from "will it happen" to "what do they get when it does." Most organizations genuinely don't know. Data sprawl, over permissioned accounts, no automated way to answer the blast radius question.
That gap was always a problem. With exploit-based breaches up 34% annually, it's now a dangerous one.
Clean Up Human and Non-Human Identity
Credential abuse is the leading entry point, and increasingly it's machine credentials rather than human ones. API keys, service accounts, OAuth tokens. Most organizations have hundreds of thousands of them. Almost none managed with least privilege by default.
An AI-assisted attacker who gets in can move through that layer in minutes. Organizations that have inventoried and rightsized their non-human identity surface are measurably harder targets.
Measure Recovery, Not Just Prevention
A 40x patch-to-weaponize gap cannot be closed by better patch management. The success metric for the security program has to change. Not "did we prevent it" but "how fast did we find it, how contained was the damage, how quickly were we back to normal."
That's not a soft reframing. It's measurable, and boards are starting to want those numbers. CISOs building programs around recovery metrics rather than prevention theater will have a much easier time over the next 18 months.
04 · Closing
A Clearing, Not an Ending
A CISO we've been talking to this past week asked us whether Mythos and Glasswing announcements felt like a watershed moment for the profession.
The assumptions that were holding things together, that defenders could keep pace with patch cycles, that perimeter-plus-endpoint was a reasonable model, those models are quickly becoming obsolete. What replaces them is: real-time exposure management, risk management that knows its own blast radius, identity governance that is comprehensive, program design built around recovery rather and prevention.
The teams and companies doing that work now aren't behind. They're right on time. The ones who get there first don't lose.
What this Means for Innovators
The infrastructure categories that define the next phase of enterprise security, exposure management, data security, non-human identity, detection and response, are where the most interesting companies are focused right now. Security leaders who understand these areas at a depth beyond marketing language will make smarter bets, give better advice, and build programs that actually hold up.
- By Pete Chronis
