# Cloudflare releases open-source Security Audit Skill for coding agents

> Cloudflare's open-source Security Audit Skill has coding agents hunt for vulnerabilities, then has separate verifier agents try to disprove each finding.

- Status: reported (One outlet, not independently verified yet.)
- Topic: Safety & Security
- Published: 2026-11-16T14:09:45+00:00
- Updated: 2026-11-16T15:10:11.845674+00:00
- Web page: https://ai-news.ikangai.com/story/cloudflare-releases-open-source-security-audit-skill-for-coding-agents/
- Seen in: AI Engineering
- Change: capability (New capability)
- Applies to: cloudflare/security-audit-skill

## What changed

Cloudflare released its open-source Security Audit Skill, which has coding agents hunt for vulnerabilities and has separate verifier agents try to disprove each finding.

Cloudflare has released its Security Audit Skill, an open-source workflow that gives coding agents a structured way to investigate possible security vulnerabilities. Hunting agents search the code for candidate issues. Each candidate then goes to a fresh verifier that did not discover it, and that verifier tries to disprove the finding before it reaches a report.

The design rests on the idea that a suspicious line of code is only the start of an investigation. A finding has to explain how a request reaches a vulnerable operation, which protection fails, and what the failure allows. An agent could flag a problem as unauthorized access, but that claim depends on facts outside the single line. The investigation has to check whether the database layer already restricts queries to the authenticated user, and whether document IDs can point to records owned by another account. It therefore has to trace where the ID enters, which user identity reaches the query, where ownership is enforced, and whether a local test can show access to another test user's document.

The verifier looks for a reason the claim cannot happen. For the document endpoint, that means inspecting the database wrapper and any access-control logic around it. The verifier must establish the source path and, where possible, reproduce the decisive result independently. A test showing that a second dummy account can retrieve the document supports a much narrower claim than a general statement that the application has broken authorization. When a runtime or deployment fact stays unknown, the finding records that gap instead of turning an assumption into a confirmed vulnerability.

The skill also tracks coverage. Verification only checks candidates the agents actually found, so an audit can examine the document endpoint thoroughly and never reach file uploads or background jobs. To address this, the skill keeps an architecture record and a coverage ledger. Hunting agents receive defined areas to examine, and coverage reviewers look for gaps in the recorded checks. The ledger also gives later runs a starting point, so agents can revisit changed source and unfinished areas rather than repeat the same searches.

Running the audit requires Node.js and a coding agent that supports parallel subagents. Users install the skill with npx skills add pointing to the cloudflare/security-audit-skill repository, then ask the agent to security audit the codebase. The output separates confirmed findings from unresolved candidates and rejected claims. It includes a detailed findings file and a separate record of work that still needs validation. Verification that executes repository code needs an isolated environment with restricted network access and permissions. Where those controls are missing, the skill records the verification blocker.

DeepSeek, in a separate item in the same newsletter, open-sourced dsh-superagent, a research prototype that manages ongoing coding tasks for a single GitHub repository. It runs on DeepSeek Harness and gives each task its own agent session and working directory.

Each accepted request becomes a tracked task assigned to a worker agent. A configured preparation command can create a Git worktree, giving the worker a separate checkout and branch. Separate sessions let several tasks move forward without sharing one conversation. Follow-up messages return to the existing task, so a clarification, such as whether a failing test reflects a regression or an intentional change, reaches the worker that already holds its earlier investigation. The system connects tasks to their GitHub conversations, including pull-request discussions. Authorization is handled separately. Allowed senders can make explicit requests, and proactive work must satisfy configured policies. Reading a related issue supplies context but does not authorize implementing everything discussed there.

For recovery, dsh-superagent preserves pending inputs and restarts interrupted assignments with a recovery note. A web panel shows tasks waiting for answers or review, along with blocked and failed work. Setup starts by cloning the dsh-superagent-prototype repository and asking the coding agent to read SETUP.md and set up dsh-superagent.

## Why it matters

Teams using coding agents for security reviews get a method for checking findings before they reach a report, which reduces false alarms. Verification that executes repository code, however, depends on sandboxing that the team must provide.

## What you can do

Before running the skill on a codebase, confirm you can provide an isolated environment with restricted network access and permissions for verification runs.

---

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it.

Cite as: IKANGAI AI News, "Cloudflare releases open-source Security Audit Skill for coding agents", 2026-11-16, https://ai-news.ikangai.com/story/cloudflare-releases-open-source-security-audit-skill-for-coding-agents/
