Project Glasswing: How AI Just Unearthed 10,000 Security Flaws in One Month
Project Glasswing: How AI Just Unearthed 10,000 Security Flaws in One Month
AI security just crossed a threshold nobody was prepared for. In the span of a single month, Anthropic's Mythos Preview model — working with about 50 partner organisations — found over ten thousand high- and critical-severity vulnerabilities across the world's most important software. That's not a typo. Ten thousand. In thirty days.
For Singapore developers, tech leaders, and anyone running production systems, this changes the calculus on software security fundamentally. The bottleneck is no longer finding bugs. It's fixing them fast enough before someone else does.
Project Glasswing: What Actually Happened
Anthropic launched Project Glasswing in April 2026 as a collaborative effort to secure critical software infrastructure before increasingly capable AI models could be turned against it. The idea was simple: give security-focused AI access to critical codebases and see what it finds.
What they found reshaped the entire conversation.
Within 30 days, Mythos Preview — Anthropic's specialised cybersecurity model — had identified over 10,000 vulnerabilities across the partners' systems. These weren't theoretical. Cloudflare alone reported finding 2,000 bugs, of which 400 were high- or critical-severity. Their verdict? The model's false positive rate was "better than human testers."
The Numbers Are Staggering
Let's put the scale in perspective:
- Cloudflare: 2,000 bugs found across 50+ critical-path repositories
- Mozilla: 271 vulnerabilities in Firefox 150 — over ten times more than what Claude Opus 4.6 found in Firefox 148
- Open-source projects: Mythos scanned 1,000+ projects and estimates 6,202 high- or critical-severity vulnerabilities. Of those already verified, 90.6% were valid (true positives)
- UK AI Security Institute: Mythos Preview is the first AI model to solve both of their cyberattack simulation ranges end to end
- Bug bounty platforms: Third-party security platform XBOW reports "absolutely unprecedented precision"
What Makes Mythos Different
Previous AI models could find bugs. Mythos Preview can chain them into working exploits.
According to Cloudflare's engineering team, the key difference is exploit chain construction. A real attack doesn't use one bug — it chains several small attack primitives together. Mythos can take multiple low-severity flaws that would normally sit invisible in a backlog and combine them into a single, severe exploit. It generates proof-of-concept code, compiles it in a sandbox, and iterates when it fails. It reasons like a senior security researcher, not an automated scanner.
Why This Matters for Singapore
Now, you might be thinking: this is a US-centric Anthropic story. What does it have to do with Singapore?
Everything — because our tech ecosystem runs on the same software.
Singapore's Heavy Open-Source Dependence
Singapore's digital economy — from Smart Nation initiatives to MAS-regulated fintech — depends heavily on open-source infrastructure. Cloudflare's infrastructure, Mozilla's Firefox, and the cryptographic libraries scanned by Mythos are the same tools that power Singapore's government portals, banking apps, and startup stacks.
Consider wolfSSL, a cryptography library used by billions of devices worldwide. Mythos constructed an exploit allowing attackers to forge SSL certificates — essentially creating fake bank or email login pages that look perfectly legitimate. The vulnerability (CVE-2026-5194) has been patched, but it illustrates the new reality: your security posture depends not just on your code, but on your entire supply chain.
The Patching Bottleneck Is Real
Project Glasswing's most sobering finding isn't technical — it's operational. Finding bugs is now the easy part. The bottleneck is triaging, verifying, and patching them.
Anthropic reports that high- or critical-severity bugs take an average of two weeks to patch. Open-source maintainers have actually asked the team to slow down disclosures because they can't keep up. Several noted they're "severely capacity constrained."
For Singapore companies running lean engineering teams — most startups and many SMEs — this creates a genuine risk. The same AI tools that defenders can use to find bugs can, in the wrong hands, find attack vectors faster than your team can patch them.
Local Implications
The Cyber Security Agency of Singapore (CSA) has been actively promoting vulnerability disclosure programmes. Project Glasswing's results suggest these programmes need to scale up dramatically — and that organisations should prepare for an influx of AI-discovered vulnerabilities.
For MAS-regulated financial institutions, the impact is even sharper. The regulatory expectation to maintain robust cybersecurity is well-established, but the speed of AI-driven vulnerability discovery may outpace traditional patch cycles. Tech leaders need to ask: when an AI finds a critical vulnerability in your payment gateway's dependency chain, how fast can you remediate?
The Pentagon, Autonomous Warfare, and AI's Ethical Crossroads
Anthropic's work with Mythos hasn't been without controversy. As The Verge reported, Anthropic's engagements with the Pentagon have highlighted the risks of autonomous warfare. The company is walking a tightrope: pushing cybersecurity forward while trying to prevent the same capabilities from enabling offensive cyber operations.
Cloudflare's team documented this tension. They found that Mythos's organic guardrails are inconsistent — the same task, framed differently, produced completely different outcomes. A model might refuse to write an exploit for one session, then produce one freely after a seemingly unrelated change. This inconsistency means safety can't be left to model behaviour alone; it requires structural safeguards.
For Singapore — which positions itself as a trusted AI hub — this raises important questions about AI governance. Singapore's Model AI Governance Framework emphasises transparency, explainability, and human oversight. Project Glasswing's results show that human oversight isn't just a nicety — it's a necessity when models can find bugs faster than humans can patch them.
What This Means for Singapore Developers
For the working developer in Singapore, three takeaways stand out:
As I covered in my guide to securing AI-powered developer toolchains, the fundamentals still matter — but the stakes are higher now.
1. Update Your Dependencies — Seriously
Mozilla patched 271 Firefox vulnerabilities. Palo Alto Networks released five times as many patches as usual. Microsoft warned that Patch Tuesday will "continue trending larger." These aren't isolated incidents — they're the new normal. If you're not keeping dependencies current, you're falling behind.
2. AI Security Tools Are Not Optional
The same models that found 10,000 vulnerabilities can also find yours. Integrating AI-powered security scanning into your CI/CD pipeline is no longer a nice-to-have. Tools like those emerging from Project Glasswing are becoming baseline requirements. If you're still relying purely on human code review for security, you're already behind.
3. Plan for a Patch Surge
Your incident response plans need to account for AI-speed vulnerability discovery. Build slack into your engineering sprints. Have a rapid response protocol for dependency patches. Consider what you'd do if a critical vulnerability is disclosed in a library your entire platform depends on.
The Bigger Picture
Project Glasswing marks a genuine inflection point. The security industry has spent decades trying to find vulnerabilities faster. AI just solved that problem. Now the question is whether the rest of the ecosystem can catch up.
As I wrote in a previous post about Singapore's AI paradox, the gap between AI capability and organisational readiness is the defining challenge of 2026. Project Glasswing makes that gap alarmingly visible. And for Singapore developers building on open-source foundations, the message is clear: the AI security revolution is here. It's not coming — it's already found 10,000 bugs in month one.
The question isn't whether AI will find vulnerabilities in your software. It's whether you'll have patched them before someone else exploits them.
Ready to secure your stack? Start by reviewing your dependency update cadence, set up automated vulnerability scanning in CI/CD, and subscribe to the CSA's cybersecurity alerts. The AI security era doesn't wait for your next sprint cycle.
Photo by Pexels | AI cybersecurity concept
