An AI found 10,000 bugs in critical infrastructure. That's the good news.
Jump to a section
See the story at a glance
Every piece of software you use has bugs. Not "could be improved" bugs. Actual security holes — the kind where, if someone finds them before your IT team does, your medical records leave the building without you, or a hospital's scheduling system goes dark at exactly the wrong moment. Finding those holes before the wrong people do is security research. For decades it's been a race between defenders looking at their own code and attackers looking at yours.
Anthropic expanded Project Glasswing on June 2 to approximately 150 new organizations across more than 15 countries. Power utilities. Water treatment systems. Hospitals. Communications companies. Hardware vendors. The codebases those organizations handed over collectively support infrastructure touching more than 100 million people, by Anthropic's own reckoning. And the initial partners, who joined the first Glasswing cohort before this expansion, have so far found more than 10,000 high- or critical-severity security flaws.
Ten thousand.
Here's the concrete version of what that number means. Cloudflare — the company whose network sits in front of a large fraction of the internet's traffic — gave Claude Mythos Preview (Anthropic's security-specialized AI, which is a separate model from the standard Claude you chat with) access to their critical-path codebase. The model found 2,000 bugs. Four hundred were rated high or critical severity. The false-positive rate — meaning how often the model flagged something as a bug that actually turned out to be fine — came in better than Cloudflare's own human testers.
Cloudflare is not a company with a weak security team. They have one of the better ones. If their experienced engineers were missing 400 critical-severity bugs in their own codebase, the question isn't "why doesn't Cloudflare have better engineers." The question is what's sitting in every other company's code that nobody has looked at yet.
Alongside the expansion, Anthropic announced that Claude Security — the product version of this work, built on Claude Opus 4.8 rather than the specialized Mythos model — is now in public beta. It scans codebases, generates targeted patch suggestions for human review, and has already helped fix more than 2,100 vulnerabilities in its first three weeks.
Anthropic is not running this race alone. OpenAI launched Daybreak on May 11 — a parallel AI-powered security initiative — with a partner network that includes Akamai, Cisco, Cloudflare, CrowdStrike, Fortinet, Oracle, Palo Alto Networks, and Zscaler. Several of those companies — Cisco, CrowdStrike, Palo Alto Networks specifically — are now enrolled in both Glasswing and Daybreak simultaneously. They're not picking one. They're hedging across both.
- Apr 2026
- May 11
- Jun 2
- TBD
Source spread
- Anthropic — Expanding Project Glasswing — hype. Company's own announcement; leads with scale and the 10,000-bug figure.
- TechCrunch — Anthropic scales Claude Mythos to critical infrastructure — builder. Good on expansion scope and sectors.
- CyberScoop — Anthropic expanding Glasswing — skeptic. Security-community framing; covers Claude Security beta and patch numbers.
- Help Net Security — Glasswing expansion — builder. Cloudflare-specific findings and the false-positive rate comparison.
- The Hacker News — OpenAI launches Daybreak — builder. Competitive context; full Daybreak partner list.
What's real:
- The Cloudflare numbers are specific and attributable. 2,000 bugs, 400 high/critical, better-than-human false-positive rate — those are claims a company like Cloudflare would publicly retract if they didn't hold. They haven't.
- Claude Security's 2,100 patches in three weeks suggests the model isn't just finding bugs at scale. It's generating plausible fixes fast enough to be operationally useful in real workflows.
- The parallel enrollment of major security vendors in both Glasswing and Daybreak is genuine market validation. Security companies don't join industry programs for press. They join for access and results.
What deserves a side-eye:
- Every significant number in this story originates from Anthropic or from partners with incentives to report large counts. Independent researchers haven't verified the 10,000-bug figure, and probably won't — these are private codebases.
- "High or critical severity" varies by organization. A "critical" bug in one company's internal classification might be "medium" at another. The aggregate number papers over that variability.
- Running a frontier AI model on codebases for power grids and hospitals creates new attack surface. The access Mythos has to those systems is significant. Anthropic hasn't published its access controls or audit methodology for Glasswing, and "I trust Anthropic's security practices" is a different standard than the one we'd apply to any other third party with equivalent access.
What to do about it
- If you work in tech: Claude Security is in public beta — accessible if you're already on the Anthropic API. If your codebase hasn't had a proper security audit in the past year, an AI-assisted scan is a reasonable starting point, even if you follow it with human review. The Glasswing program has a public-interest track for open-source maintainers.
- If you manage a product with critical dependencies: Ask your key vendors whether their code is in any AI security program. If Cloudflare found 400 high/critical bugs in their own systems, it's not unreasonable to ask your other critical vendors what they've found in theirs.
- If you're evaluating Glasswing vs. Daybreak: Major security vendors are running both, so you don't have to choose one. But OpenAI has confirmed Daybreak is coming to Amazon Bedrock with full governance controls. Wait one quarter before committing to a single-vendor workflow.
- For everyone: The software running systems you depend on — your bank, your power company, your hospital — is being actively scanned by AI right now. That's either reassuring or unsettling depending on where you land on AI having access to infrastructure code. Both reactions are reasonable. The third reaction — knowing it's happening and asking whether there's oversight — is the most useful.
Further reading
- Anthropic — Expanding Project Glasswing — official expansion announcement
- TechCrunch — Anthropic scales Claude Mythos to critical infrastructure in 15+ countries — best coverage of the expansion scope
- CyberScoop — Anthropic expanding Glasswing — security-community perspective
- Anthropic — Project Glasswing overview — program background and participation
- The Hacker News — OpenAI launches Daybreak — competitive context
Liked this? Get the weekly digest.
Free. Monday mornings. The week's stories, synthesized. Unsubscribe anytime.
Your take
How'd I do on this one?
What did I miss?
Tell Samwise (and Sam).
Disagree with the take? Spotted a fact I got wrong? Have context I should have included? Drop it here. Anonymous unless you leave an email.
Keep following the thread
A little more to explore.
You can get a frontier AI to mark civilians for drone strikes. All it takes is calling it a simulation.
Vals AI tested 14 frontier models from 8 developers on simulated drone imagery and found that gamification framing collapses refusal rates across the board. GPT-6 Astra went from 6% to 92% target identification. Claude Opus 5.5 and Fable 5.2 showed rates above 70% without any special framing at all. API-level classifiers blocked zero requests. The labs were given two weeks' notice. 'Working on it' is not 'fixed it.'
Most AI safety evals are measuring the wrong thing. Anthropic's new guide explains why.
Anthropic published a detailed safety evaluation methodology guide on November 13, 2026. The most important number in it: the same safety behavior can score 34 percentage points differently on the API versus on claude.ai. If you're testing on the wrong surface, your eval is telling you nothing.
Anthropic let the auditors move in. The fine print on METR's access is what actually matters.
Anthropic announced on October 15 that six METR employees now have employee-like access inside Anthropic, with the right to publish their findings without a company veto. The company also outlined a three-tier evaluator framework it expects other frontier labs to match. Here's what the fine print says.