Project Glasswing and the New Baseline for Cybersecurity Assessment

Anthropic's Project Glasswing, announced on April 7, 2026, deploys an unreleased AI model to find and patch vulnerabilities across the world's most critical software infrastructure. The initiative brings together Amazon, Apple, Microsoft, Google, CrowdStrike, Palo Alto Networks, and others, backed by $100 million in usage credits and $4 million in direct donations to open-source security organizations. For professional services firms that advise on cybersecurity risk, technology due diligence, or IT posture assessments, the announcement resets the baseline for what a competent assessment looks like. For the underlying technical context on the model powering Glasswing, see our companion analysis of Claude Mythos Preview and the decision not to release it.

Key Takeaways

• Claude Mythos Preview found vulnerabilities that survived decades of human review and millions of automated security tests, including flaws in every major operating system and browser
• The Linux Foundation's CEO described the initiative as enabling AI-augmented security to become accessible to maintainers who previously could not afford dedicated security teams
• CrowdStrike, Palo Alto Networks, and Microsoft, companies with their own proprietary AI security tools, publicly endorsed Anthropic's model as superior for vulnerability discovery
• The 12 launch partners plus 40 additional organizations with access represent significant portions of global software infrastructure
• The capability gap closed faster than forecast: Claude Fable 5 (9 June 2026) brought Mythos-class capability to general availability with cyber safeguards, nine weeks after the Glasswing launch
• Organizations relying on conventional vulnerability assessments completed before April 2026 are now benchmarked against a demonstrably lower standard

Scroll right to see more

Scroll right to see more

The Gap Between Current Practice and Current Capability

Most cybersecurity assessments delivered by professional services firms follow a well-established methodology: automated vulnerability scanning, manual penetration testing, configuration review, and compliance mapping against frameworks like ISO 27001, NIST CSF, or SOC 2. These assessments are competent for the threat landscape they were designed to address.

The problem is that the threat landscape just shifted.

Claude Mythos Preview identified a line of vulnerable code in FFmpeg (one of the most widely used video processing libraries in the world) that automated testing tools had executed five million times without catching the issue. The vulnerability had existed for 16 years. In OpenBSD, a system specifically engineered for security, the model found a flaw that had been present for 27 years. These are not esoteric edge cases. These are production systems running in enterprise environments today.

The model did not require human guidance for most of these discoveries. It found and reported vulnerabilities autonomously, including chaining multiple Linux kernel flaws together to achieve full system control. External testers confirmed that it completed end-to-end corporate network attack simulations that would take a skilled human over 10 hours.

When CrowdStrike, Palo Alto Networks, and Microsoft (companies that have built their businesses on proprietary AI-powered security) publicly endorse a competitor's model as the standard for vulnerability discovery, the signal is clear. The current generation of security assessment tools and methodologies, including those sold by the endorsing companies, has a ceiling that AI has moved past.

What This Means for Advisory Firms

Three practical implications apply to any firm delivering cybersecurity advisory, technology due diligence, or IT risk assessments.

Assessment scope needs to expand beyond known vulnerability databases. Conventional scans check against databases of known vulnerabilities (CVEs). The vulnerabilities that Mythos Preview found were zero-days. They did not exist in any database. A scan that reports "no critical vulnerabilities found" against CVE databases tells you nothing about whether zero-day exposure exists. This distinction matters in every engagement where a client relies on assessment results to make investment, insurance, or compliance decisions. For M&A deal teams, this issue now sits at the center of the cybersecurity due diligence framework rather than at the periphery.

The definition of "reasonable security measures" is shifting. Regulatory frameworks and industry standards generally require organizations to implement "reasonable" or "proportionate" security measures. What qualifies as reasonable is benchmarked against prevailing practices. When AI systems can identify vulnerabilities that entire security teams and automated tools have missed for decades, the prevailing-practice benchmark moves. Organizations that could demonstrate reasonable care last quarter may face harder questions next quarter, not because they did anything wrong, but because the definition of adequate diligence evolved. The four-framework regulatory alignment across NIS2, DORA, CRA, and the revised CSA becomes more complex when the underlying "state of the art" reference point shifts mid-assessment cycle.

Due diligence reports need a capability disclaimer. Any cybersecurity assessment or technology due diligence report delivered after today should address whether AI-augmented vulnerability discovery was used, or explicitly state that it was not. Acquirers, investors, and boards who rely on these reports deserve to understand the methodology's limitations relative to what is now technically possible. This is not a marketing pitch for new tools. It is a disclosure obligation for anyone providing professional opinions on security posture.

The Open-Source Dimension

The Linux Foundation's involvement in Project Glasswing highlights a structural vulnerability in the software ecosystem. Jim Zemlin, the Foundation's CEO, was direct: security expertise has historically been a luxury available to organizations with large security teams, while open-source maintainers (whose software underpins most of the world's critical infrastructure) have been left to handle security independently.

This matters for advisory work because virtually every enterprise technology stack depends on open-source components. A client's security posture is only as strong as the weakest link in its dependency chain. If the open-source libraries embedded in a client's systems contain undiscovered zero-days, and Anthropic's results suggest many do, then assessments that stop at the client's proprietary code boundary are incomplete by design. This ties directly into the bidirectional supply chain risk AI development tools create, where dependency compromise is already a primary attack vector.

Anthropic's $4 million donation to open-source security organizations through the Linux Foundation, including $2.5 million to Alpha-Omega and OpenSSF and $1.5 million to the Apache Software Foundation, is a starting point. But the scale of the problem (undiscovered vulnerabilities in software running billions of devices) requires more than donations. It requires a structural change in how open-source security is funded, assessed, and maintained.

The Patching-Gap Reality

A baseline shift in discovery is only half the story. The other half is what happens to those findings once they exist. Reporting through April 2026 indicates that fewer than 1% of the vulnerabilities surfaced through Project Glasswing / Claude Mythos Preview have been patched — even with responsible disclosure to vendors. The bottleneck is not finding the bugs. It is the engineering capacity, vendor cooperation, and patch-distribution channels that absorb them downstream.

For assessment methodology, the implication is sharp. A current assessment that maps an organization's known-CVE exposure is incomplete in a world where AI is finding vulnerabilities faster than the ecosystem can ship fixes. The relevant question becomes "how exposed is this client to the disclosed-but-unpatched category?" — not just to the published-CVE backlog. That is a different scoping conversation, with different remediation playbooks (compensating controls, vendor pressure, depend-on-it inventories).

The Timeline Question

At launch, Anthropic stated that it did not plan to make Claude Mythos Preview generally available, but that its eventual goal was to enable users to deploy models with these capabilities at scale, with appropriate safeguards. This article's original estimate was that the capability gap between Project Glasswing participants and the rest of the market would close within 12 to 18 months, possibly sooner.

It closed in nine weeks. On 9 June 2026, Anthropic released Claude Fable 5, the first publicly available Mythos-class model, with cybersecurity-sensitive queries classifier-routed to the older Claude Opus 4.8 (Anthropic reports more than 95% of sessions involve no fallback, and zero compliance on harmful single-turn cyberattack requests across 30 public jailbreak techniques). Project Glasswing participants were upgraded to the unsafeguarded Claude Mythos 5, alongside a US-government collaboration. AI-powered vulnerability discovery at the general-capability level is now one API call away for any organization; the unrestricted cyber variant remains gated to the defensive consortium.

The advisory-firm implication sharpened rather than changed. The window to be "ahead of the curve" did not last 12 months; it lasted one fiscal quarter. Firms that began building AI-augmented assessment methodologies in April have tested workflows and documented case studies today. Firms that waited are now operating in the environment this article anticipated, without the runway. The same dynamic we covered in our PE firm's guide to cybersecurity due diligence applies with the timeline compressed: the cost of adoption is far lower than the cost of being the firm that delivered the clean assessment before the AI-augmented audit found the problems. For the executive metrics that follow from machine-speed discovery (the Velocity Gap and the Blast Radius Index), see the Velocity Gap doctrine.

Practical Steps

For firms advising on cybersecurity, technology risk, or IT due diligence, four actions apply immediately.

First, review how your current assessment methodology accounts for zero-day exposure. If it does not, document that limitation and communicate it to clients. Second, evaluate whether your tooling pipeline can integrate AI-augmented vulnerability discovery when it becomes broadly available. Anthropic has indicated this is a matter of when, not whether. Third, update your engagement scoping to address open-source dependency analysis. If your assessments do not map the client's open-source supply chain, you are leaving a known gap. Fourth, monitor the Cyber Verification Program that Anthropic intends to launch. Early access to AI-powered security tools will differentiate firms that move first.

The full Intelligence Brief covers the detailed Project Glasswing partner analysis, AI-augmented assessment methodology frameworks, open-source dependency risk mapping, and a comparative timeline for when these capabilities become broadly accessible.