AI Code Audits Hit New Milestone: Anthropic and Mozilla Show Zero-Day Detection at Scale

AI-driven code review has crossed a critical threshold. Until early 2026, the consensus held that large language models could spot known antipatterns but lacked the understanding to find truly novel vulnerabilities in well-audited codebases. That line has now been erased by a series of high-impact disclosures from Anthropic and Mozilla, detailed in a recent Heise+ report.

The practical consequence for security teams is immediate: the weakest link is no longer the detection tool but the human capacity to triage the flood of findings. Organisations that rely on manual code review must adapt their workflows or risk being overwhelmed.

New capabilities, new discoveries

In February 2026, Anthropic’s Frontier Red Team published a report describing how Claude Opus 4.6, without any specially tuned setup, identified over 500 previously unknown high-severity vulnerabilities across major open-source projects. The scale of the find was unprecedented for an AI model.

Two months later, Claude Mythos Preview went further. According to Anthropic, the model discovered zero-day vulnerabilities in every major operating system and every major browser. One of the most striking findings was a 17-year-old unauthenticated remote code execution flaw in FreeBSD, tracked as CVE-2026-4747. The bug had survived decades of human auditing and conventional fuzzing.

Mozilla confirms the trend

Mozilla’s own experience, detailed in a Hacks blog post on May 7, 2026, provides the most concrete proof of the new reality. The organisation deployed an agentic harness built on Claude Mythos Preview to scan Firefox’s codebase. The result: Firefox 150 patched 271 latent security bugs, including multiple sandbox escapes that had eluded years of fuzzing.

Overall, Mozilla fixed 423 security bugs in April 2026. That is a dramatic leap from the typical 20–30 monthly fixes the organisation had previously averaged. In its post, Mozilla described the effect succinctly: “Suddenly, the bugs are very good.” For defenders, the news is double-edged — detection has improved, but the operational burden has multiplied.

Key facts

The bottleneck shifts to human review

The new detection capability has created a fresh bottleneck. The Heise report notes that the scarcest resource in the entire security process is now the human reviewer who must assess and prioritise the AI-generated findings. Automated tools can surface hundreds of potential issues, but each one requires expert judgment to confirm, reproduce, and remediate.

Security teams that have not yet integrated AI-assisted triage pipelines risk falling behind. The model itself is not the limiting factor; the organisational ability to act on its output is. Some firms are beginning to experiment with agentic workflows that automatically classify and assign severity, but the process remains largely manual.

Limits and next checks

While the results are impressive, several caveats remain. The Anthropic reports come from the vendor itself, and independent verification of the full set of findings is still pending. Mozilla’s numbers are specific to its own codebase and may not generalise to every organisation. The models also require significant compute resources, and their effectiveness depends on the quality of the prompt and context provided.

For security teams, the immediate next step should be to test AI-assisted code review in a controlled environment, compare its output against existing tooling, and begin planning for a triage workflow that can handle an order-of-magnitude increase in findings. The era of AI discovery is here; the era of scalable human response is still being built.

Source: Heise+ (paywalled) – “Security: Wie Sie KI-Codereviews sinnvoll nutzen” (July 30, 2026)