Claude Found 22 Firefox Zero-Days in Two Weeks — and That Changes EverythingClaude 两周找到 22 个 Firefox 零日漏洞——游戏规则变了

The Setup

Anthropic partnered with Mozilla to stress-test Firefox — one of the most well-tested open-source projects in the world. They pointed Claude Opus 4.6 at the codebase for two weeks. The result: 22 vulnerabilities discovered, 14 classified as high-severity. That’s nearly a fifth of all high-severity Firefox bugs remediated in 2025 — found in 14 days.

The first Use After Free vulnerability was flagged within twenty minutes of exploration. By the time human researchers finished validating that first bug, Claude had already found fifty more unique crashing inputs.

Key Takeaways

• Finding is cheap, exploiting is hard. Claude spent $4,000 in API credits trying to turn its discoveries into working exploits — and only succeeded twice, in a sandbox-disabled test environment. The asymmetry is massive: finding bugs costs pennies per vulnerability, exploiting them costs orders of magnitude more.

• Task verifiers are the unlock. Claude’s best results came when it could check its own work with external tools — confirming bugs were real, patches actually fixed them, and nothing else broke. This “agent + verifier” pattern is becoming the standard architecture for AI-powered code work.

• Scale is the story. Claude scanned nearly 6,000 C++ files and submitted 112 unique reports. No human team operates at that throughput. And Mozilla’s response — encouraging bulk submission without manual pre-validation — signals that maintainers are adapting to AI-speed workflows.

Why It Matters

This isn’t a benchmark result or a demo. It’s a production collaboration with real CVEs fixed in Firefox 148, shipped to hundreds of millions of users.

For builders: the “agent + task verifier” pattern Anthropic describes is the same architecture that makes coding agents useful. The model proposes, the verifier validates, the agent iterates. If you’re building AI tools, this is the loop to study.

For investors: AI security tooling is about to become a massive category. The asymmetry — cheap to find, expensive to exploit — is a net positive for defenders, but only if they adopt fast enough.

What to watch:

• Whether other frontier models (GPT-5.4, Gemini) replicate these results on comparable codebases
• Mozilla’s internal adoption of Claude for ongoing security — first customer or proof-of-concept?
• The exploit gap closing as models improve — Anthropic flagged this as a concern worth monitoring