Anthropic's Project Glasswing Found a 27-Year-Old Bug in OpenBSD — Before Hackers Could

The Facts

Anthropic launched Project Glasswing as a controlled initiative giving select organisations access to Claude Mythos Preview — an unreleased frontier model with significantly enhanced vulnerability discovery capabilities. The project represents the first formal deployment of an AI system specifically designed for offensive security research at scale, with appropriate controls to prevent dual-use exploitation.

The results from the initial weeks of internal testing are striking. Mythos Preview identified thousands of zero-day vulnerabilities across every major operating system and web browser. Among the discoveries was a 27-year-old vulnerability in OpenBSD — a security-focused operating system widely used in network infrastructure globally. A bug of that age, in a codebase explicitly written with security as a primary objective, demonstrates that even mature, carefully maintained codebases contain classes of vulnerability that human security researchers have consistently missed.

Anthropic committed over $100 million in model usage credits to the initiative. The company has explicitly stated it has no plans for public release of Mythos due to its dual-use cybersecurity risks — the same capability that enables finding and patching vulnerabilities also enables their weaponisation. This represents one of the clearest cases in recent AI development history where capability advance has required explicit containment rather than open release.

Technical Deep-Dive

The technical capability enabling Mythos Preview's vulnerability discovery at scale is likely a combination of enhanced code comprehension and novel reasoning about security-relevant program states. Current frontier models can read and understand large codebases — but converting that understanding into systematic vulnerability discovery requires reasoning about complex interactions between code components, memory states, and execution paths that can be triggered by adversarial inputs.

The discovery of a 27-year-old vulnerability in OpenBSD is particularly illuminating. Security researchers have manually reviewed OpenBSD code for decades, with the project having a strong culture of security-focused development and peer review. The vulnerability survived because human code review is necessarily sampling-based — reviewers check the most suspicious or recently changed code, but cannot achieve complete coverage of a large codebase. AI-assisted code analysis can achieve coverage that human review cannot, systematically examining every code path for classes of vulnerability that require pattern recognition across thousands of related code instances.

The controlled deployment model — giving access to select partner organisations rather than releasing the tool publicly — reflects the asymmetry between offensive and defensive use. A defender needs to find and patch all vulnerabilities; an attacker needs to find only one. A publicly available AI tool that generates zero-day exploits at scale would dramatically advantage attackers over defenders.

The ASEAN Perspective

For cybersecurity teams in Singapore, Malaysia, and Indonesia, Project Glasswing has implications that extend beyond the specific vulnerability disclosures. The initiative demonstrates that AI-powered vulnerability discovery is technically feasible at scale — and that the organisations with access to the most capable AI systems will develop a significant advantage in identifying and patching vulnerabilities in their infrastructure before adversaries find them.

Singapore's CSIRT (Computer Emergency Response Team) and the Cyber Security Agency's Government Technology Agency partnerships position Singapore to benefit from disclosures through responsible vulnerability programs. ASEAN enterprises using affected software (every major OS and web browser) should monitor security advisories from their software vendors as Glasswing findings are disclosed through coordinated vulnerability disclosure processes.

The dual-use concern is particularly acute for ASEAN's regulatory environment. Singapore, Malaysia, and Indonesia all have computer misuse legislation that criminalises unauthorised access to computer systems — but the legislation was not written with AI-powered vulnerability discovery in mind. The legal frameworks for responsible AI-assisted security research across ASEAN are largely undefined.

RECATOOLS Verdict

Project Glasswing represents the responsible end of the spectrum for deploying powerful AI security capabilities. Anthropic's decision to operate through a controlled programme with select partners rather than public release reflects appropriate consideration of the dual-use risk — even at the cost of significantly slower vulnerability remediation across the broader software ecosystem.

For ASEAN enterprise security teams, the practical implication is to maintain rigorous patch management for major OS and browser updates. The vulnerability disclosures flowing from Project Glasswing will likely accelerate the pace of security patches from major vendors over the next 12-24 months. Staying current on patches is now more important than ever, as AI-powered discovery has demonstrably shortened the window between vulnerability existence and potential exploitation.

Frequently Asked Questions