Anthropic just announced Project Glasswing, and if you're a CISO, CTO, or VP of Engineering, you need to read this. They've built an AI model—Claude Mythos Preview—that autonomously finds and exploits zero-day vulnerabilities in every major operating system and browser. Not theoretical bugs. Not toy examples. Critical, exploitable vulnerabilities that have survived decades of human review and millions of automated security tests.
This isn't a research demo. It's a watershed moment for enterprise cybersecurity.
What Anthropic Just Did
Anthropic trained Claude Mythos Preview as a general-purpose AI model. They didn't explicitly train it for cybersecurity—these capabilities emerged as a downstream consequence of improvements in code, reasoning, and autonomy. The same improvements that make it better at patching vulnerabilities also make it better at exploiting them.
Here's what Mythos Preview found autonomously (without human intervention):
- A 27-year-old vulnerability in OpenBSD (one of the most security-hardened operating systems in the world) that allows remote denial-of-service attacks.
- A 16-year-old FFmpeg vulnerability in a line of code that fuzzers had hit 5 million times without catching the problem.
- Thousands of high- and critical-severity zero-days in every major OS (Linux, FreeBSD, macOS, Windows) and every major browser (Chrome, Firefox, Safari, Edge).
- Remote code execution exploits that chain together multiple vulnerabilities to grant full root access.
- JIT heap spray exploits in web browsers that escape both renderer and OS sandboxes.
In one case, Mythos Preview chained together four vulnerabilities in a single exploit. In another, it wrote a 20-gadget ROP chain split across multiple network packets to gain root access to a FreeBSD NFS server.
The Numbers That Should Wake Up Every CISO
- Cost to find the OpenBSD vulnerability: Under $50 (one run).
- Total cost for 1,000 runs across the entire codebase: Under $20,000.
- Validation rate: 89% of manually reviewed bug reports matched the model's severity assessment exactly; 98% were within one severity level.
- False positive rate: Near zero. (Memory-unsafe bugs are verified with Address Sanitizer, which perfectly separates real bugs from hallucinations.)
Bottom line: Anthropic found thousands of critical vulnerabilities for the cost of a single penetration test.
Why This Changes the Cybersecurity Playbook
1. The Window Between Discovery and Exploitation Just Collapsed
CrowdStrike's George Kurtz put it bluntly: "The window between a vulnerability being discovered and being exploited by an adversary has collapsed—what once took months now happens in minutes with AI."
If you're a CISO, this is your new threat model:
- Attackers will use AI models to find zero-days in your systems faster than your team can patch them.
- N-day vulnerabilities (known but unpatched) will be weaponized within hours of disclosure.
- The cost of offensive capabilities just dropped to near-zero for well-resourced attackers.
2. Defense-in-Depth Just Got Harder
Mythos Preview demonstrated the ability to chain together vulnerabilities that individually might not be exploitable. For example:
- It bypassed KASLR (kernel address space layout randomization) by chaining a read vulnerability with a write vulnerability.
- It turned a one-bit write into a full privilege escalation by finding adjacent memory corruption bugs.
Translation for business leaders: Security mitigations that rely on "friction" (making exploitation tedious) are less effective against AI. Mitigations that impose hard barriers (like W^X or hardware-enforced isolation) remain critical.
3. Open Source Maintainers Are Outgunned
Anthropic found vulnerabilities in the most audited codebases in the world:
- FFmpeg (used by nearly every video service)
- Linux kernel (runs most of the world's servers)
- OpenBSD (known for its security-first culture)
If these projects—despite millions of hours of human review and automated fuzzing—have thousands of undiscovered vulnerabilities, what about your codebase?
What Anthropic Is Doing About It
Project Glasswing: A $100M Commitment
Anthropic is giving limited access to Claude Mythos Preview to:
- 12 launch partners (including AWS, Microsoft, Google, Apple, NVIDIA, CrowdStrike, Palo Alto Networks, JPMorganChase, and Cisco).
- 40+ additional organizations that build or maintain critical software infrastructure.
- Open-source maintainers (via the Claude for Open Source program).
The goal: Give defenders a head start before models with similar capabilities become broadly available.
Key commitments:
- $100M in model usage credits for defensive security work.
- $4M in direct donations to open-source security organizations (Alpha-Omega, OpenSSF, Apache Software Foundation).
- Coordinated vulnerability disclosure (90 + 45 day timeline).
What Partners Are Saying
AWS:
"Our teams analyze over 400 trillion network flows every day for threats, and AI is central to our ability to defend at scale. Claude Mythos Preview is helping us strengthen our code before threats emerge."
Microsoft:
"When tested against CTI-REALM, our open-source security benchmark, Claude Mythos Preview showed substantial improvements compared to previous models."
CrowdStrike:
"If you want to deploy AI, you need security. That is why CrowdStrike is part of this effort from day one."
JPMorganChase:
"Promoting the cybersecurity and resiliency of the financial system is central to JPMorganChase's mission, and we believe the industry is strongest when leading institutions work together on shared challenges."
What This Means for Your Organization
If You're a CISO or VP of Security:
- Assume AI-augmented attackers are already here. Your threat model needs to account for adversaries with near-zero-cost vulnerability discovery.
- Prioritize patching. The time window between disclosure and weaponization is shrinking. Automated patch management is no longer optional.
- Invest in defense-in-depth with hard barriers. KASLR, stack canaries, and W^X are table stakes. Hardware-enforced isolation (e.g., confidential computing) is the next frontier.
- Audit your third-party dependencies. If critical open-source projects have thousands of undiscovered bugs, your supply chain is a liability.
If You're a CTO or VP of Engineering:
- Adopt AI-powered code review. Mythos Preview's capabilities will eventually be commoditized. Defensive tools will follow.
- Shift left on security. Fix vulnerabilities before code ships. AI models can now scan for bugs faster than humans can write them.
- Plan for AI-assisted penetration testing. Your red team should be using these tools. If they're not, your adversaries will be.
If You're a CFO or COO:
- Cybersecurity budgets need to reflect the new threat landscape. The cost of a breach just went up. The cost of prevention needs to match.
- Insurance policies may not cover AI-augmented attacks. Review your cyber insurance terms. Exclusions for "acts of war" or "state-sponsored attacks" may apply.
- Compliance regimes are lagging. SOC 2, ISO 27001, and NIST don't yet account for AI-powered attackers. Expect regulatory changes.
The Broader Implications
1. This Is the New Normal
Anthropic is not releasing Claude Mythos Preview publicly. But the capabilities it demonstrates are not unique. OpenAI, Google, and other frontier labs are building similar models. Within months, these capabilities will proliferate.
2. Defenders Have a Short-Term Advantage (If They Move Fast)
Project Glasswing gives defenders a 90-day head start to find and fix vulnerabilities before attackers get access to similar tools. But only if they act now.
3. The Industry Needs Standards
Anthropic is calling for collaboration on:
- Vulnerability disclosure processes (how to handle AI-discovered bugs at scale).
- Software development lifecycle standards (secure-by-design practices for the AI era).
- Patching automation (because humans can't keep up).
An independent, third-party body—combining private and public sector organizations—may be the ideal home for this work.
What You Should Do This Week
- Read Anthropic's full announcement at anthropic.com/glasswing.
- Review your vulnerability management process. Are you patching within 90 days? 45 days? 24 hours?
- Talk to your security vendors. Ask if they're using AI-powered tools for vulnerability discovery. If not, ask why.
- Engage with your board. This is a board-level risk. Cybersecurity is no longer just an IT issue—it's a business continuity issue.
The Bottom Line
Anthropic's Claude Mythos Preview is not a warning shot. It's a proof of concept that AI models can now find and exploit vulnerabilities faster, cheaper, and more thoroughly than any human team.
For CISOs and IT leaders, this is a call to action. The window to get ahead of this threat is measured in months, not years.
For business leaders, this is a wake-up call. Cybersecurity is no longer about compliance checklists. It's about survival in an AI-driven threat landscape.
The question isn't whether AI will reshape cybersecurity. It's whether your organization will be ready when it does.
Sources:
- Anthropic: Project Glasswing
- Anthropic Red Team: Claude Mythos Preview Technical Details
- Claude Mythos Preview System Card
- Microsoft Research: Strengthening Secure Software at Global Scale
- OpenAI: Strengthening Cyber Resilience
Tools mentioned: Claude (by Anthropic), Address Sanitizer, Syzkaller, AFL (American Fuzzy Lop)
Want to calculate your own AI ROI? Try our AI ROI Calculator — takes 60 seconds and shows projected savings, payback period, and 3-year ROI.
