----------------
🛠️ Tool: security-audit, the coding-agent skill that seeded Cloudflare's vulnerability discovery harness
===================
Cloudflare has published security-audit, a coding-agent skill that turns an agent into a security auditor. Per its README, this is the single-repo starting point that grew into the multi-stage, fleet-wide vulnerability discovery harness documented in the blog post "Build your own vulnerability harness".
What the skill does
The audit runs in six phases:
1. Reconnaissance: map architecture, trust boundaries, input surfaces, prior evidence, and deterministic coverage into architecture.md and coverage-ledger.json.
2. Coverage-led hunting: isolated hunter agents are assigned from ledger units, log their checks, and coverage critics look for gaps.
3. Candidate validation: every unique candidate goes to a fresh verifier whose stated goal is to disprove it.
4. Structured output: confirmed, needs_validation, and rejected records are written to findings.json and checked against report-schema.json.
5. Independent record verification: fresh agents verify final source claims; material replacements get another independent verifier.
6. Target-neutral reporting: REPORT.md, FINDINGS-DETAIL.md, and NEEDS-VALIDATION.md derive from verified records and the coverage ledger.
Two validators run mechanically: validate-coverage-ledger.cjs fires after ledger creation and after every later ledger update, and validate-findings.cjs runs in phase 4 and again after every phase 5 replacement.
Verdict semantics
The three verdicts are asymmetric. Confirmed requires a complete source trace plus a bounded observed result. needs_validation holds one exact unresolved fact and carries no severity. Rejected records a disproved candidate. Keeping unproven candidates in a separate class, each with an exact open fact attached, is the part of the design most agent audits skip.
Additive runs
Multiple runs against the same repo are additive. The skill reuses prior ledgers and findings to target gaps, revalidates changed source, and carries forward current-source evidence without treating stale or unresolved work as covered.
Repository layout
SKILL.md covers setup, core principles, and audit anti-patterns. RECONNAISSANCE.md and HUNTING.md hold the phase 1-2 prompts. ATTACK-CLASSES.md defines core, wildcard, and obvious-things attack classes, extended by per-surface files:
• AI-AND-LLM.md: prompt injection, agent/tool abuse, output handling for LLM-backed targets
• MEMORY-SAFETY-AND-BINARY.md: memory-safety, binary, and kernel classes for native targets
• WEB-PROTOCOL-AND-AUTH.md: HTTP request framing, cache, and auth protocol classes
• CLIENT-SIDE.md: DOM injection, messaging trust, UI redress, prototype pollution
• SUPPLY-CHAIN-AND-RELEASE.md: dependencies, CI, release, signing, updates, plugins, extensions
• CLOUD-AND-DEPLOYMENT.md: IAM, IaC, containers, serverless, ingress, runtime configuration
• PROTOCOLS-RPC-AND-MESSAGING.md: RPC, serialization, queues, brokers, webhooks, streaming
• RESOURCE-EXHAUSTION-AND-AVAILABILITY.md: shared resources, quotas, workers, operator spend
• DATA-ISOLATION-AND-LIFECYCLE.md: tenant isolation, cache, search, export, backup, migration, deletion, restore
• DESKTOP-MOBILE-AND-LOCAL-IPC.md: native apps, deep links, webviews, helpers, daemons, local IPC
VALIDATION-AND-REPORTING.md covers phases 3-6, and report-schema.json defines the output format.
Use cases
• A reproducible single-repo audit pass for coding agents that support skills
• A documented base to fork when building a larger discovery harness
• Schema-validated findings that stay comparable across runs and repos
Limitations
• The README describes design, not measured outcomes; no benchmarks are attached.
• Output quality tracks the underlying coding agent and model.
• Not personally tested; details here are author-reported from the README and the Cloudflare post.
References: Cloudflare's blog post "Build your own vulnerability harness".
🔹 tool #Cloudflare #AppSec #AIAgents #vulnerability_discovery
🔗 Source: https://github.com/cloudflare/security-audit-skill
ThreatAftYesWiki 9.1: A Backtick in the Wrong Escape | ThreatAftCVE-2026-116207: unauth SQL injection in YesWiki's Bazar search API. 9.1 v3.1 / 8.8 v4.0, read-only and in-band. Fixed in 4.7.0 and 5.0.0-alpha4.
