Enterprise Security in the Agentic AI Era

Cloudflare

2 min read

Cloudflare

Sources: Cloudflare (homepage) · Cyber Frontier Models blog post

Cloudflare is an internet-infrastructure and security company and a partner in Project Glasswing. Its inclusion among the ~50 Glasswing partners places it in the coalition’s extended membership beyond the twelve named launch partners.

Glasswing Result (One Month In)

Per Anthropic’s one-month update and Cloudflare’s own Cyber Frontier Models post, Cloudflare used Claude Mythos Preview to find 2,000 bugs across its critical-path systems, of which 400 were high- or critical-severity, at a false-positive rate the Cloudflare team considers better than human testers.

FP rate “better than human testers”. Cloudflare’s false-positive-rate claim is one of the clearest first-party defender statements that frontier-AI vulnerability discovery has crossed the precision threshold where it competes with skilled human review on both recall and precision. It corroborates the harness-over-model precision argument from a deploying enterprise rather than a model vendor.

Open-Source Security Methodology

Cloudflare also publishes a security-audit methodology as open source. `security-audit-skill` is an MIT-licensed six-phase multi-agent audit with adversarial validation, distributed as a skill a coding agent installs, and Semgrep’s July 2026 survey reports it at ~2K stars.1 The phases run reconnaissance, a parallel hunt across attack classes, adversarial validation in which separate agents attempt to disprove each finding, a report, a schema-validated findings file, and an independent verification pass against source; the skill carries the methodology and the host agent supplies the model. Semgrep recommends it to a team already working inside a coding agent that wants a rigorous audit method with no new infrastructure.

See Also

Notes

Footnotes

  1. Semgrep — Comparing open source AI code security harnesses, July 2026 (no day-level date exposed; author not named). The licence, star count and six-phase framing are human-written; the per-phase detail is from Semgrep’s LLM-generated repository summary. Summarized at OSS AI Security Harness Comparison. ↩

Graph View

Backlinks