Skip to content

Tool Shortlist

Start with a tool that fits the job, then use Choosing a model to select its model and hosting route. The tools below have freely available code or prompts; model usage and infrastructure may still cost money.

The first review tutorial uses OpenCode with security-review-repo, the playbook’s adaptation of Anthropic’s security-review prompt. The comparisons below help you decide whether that approach fits your review.

Try Defending Code for a structured scan and triage workflow, or Google Mantis for a modular review pipeline. Either can replace the review skill in a second discovery pass. The specialist options below need more setup.

Repository and prompt · MIT

Discovery PR review

A lightweight security review prompt, also shipped as Claude Code’s /security-review command and a GitHub Action. It examines code changes and filters findings by confidence. Choose it for a small, inspectable starting point or a baseline to compare with larger tools.

  • Setup: the playbook supplies security-review-repo, a whole-repository adaptation of the prompt. Follow the installation and review instructions, or the full tutorial if you also need sandbox setup. The upstream command reviews pending changes; use the adaptation for an imported repository.
  • Maturity: an established prompt with a small setup burden. Evaluate the GitHub Action separately before relying on it in CI: its repository has had no maintainer commits or replies since February 2026, and issues filed in September report that newer model names fail because the Action installs an older Claude Code, and that a retired model ID silently disables its false-positive filter. The prompt itself is unaffected.
  • Limits: the supplied prompt excludes classes such as DoS, rate limiting, outdated dependencies and memory-safety issues in languages it treats as memory safe. Read these exclusions before assessing coverage. Its confidence filter can hide real findings, and it does not require reproduction.

Anthropic Defending Code Reference Harness

Section titled “Anthropic Defending Code Reference Harness”

Repository · Apache-2.0

Threat modeling Discovery Triage Patching

Skills for each review stage, plus an autonomous find → verify → patch pipeline. Choose it when you want explicit stages and structured reports. The full pipeline starts with C/C++ memory bugs, Docker and sanitizers; other stacks need adaptation.

  • Setup: start with the skills and /vuln-scan. The installation, scan and export instructions cover skill setup and report export. Importing skills does not install the autonomous pipeline or its runtime.
  • Maturity: Anthropic explicitly labels the repository unmaintained. Treat it as a reference implementation that your team will need to own.
  • Operating limits: set a budget, monitor the run and check the report before acting on its findings. Completing a scan does not establish that the findings are correct or the instructions were followed.

Repository · Apache-2.0

Discovery Reproduction Patching

A toolkit of review skills with a reference harness built on Google’s Agent Development Kit (ADK). Its stages map the code and threats, explore possible bugs, check and merge findings, then reproduce and patch them. Choose it when you want to inspect or adapt individual stages of a broader review.

  • Setup: import a pinned skill pack using the Mantis installation instructions. That section also explains which stages fit this sandbox. Running the upstream reference harness requires its own setup and model configuration.
  • Maturity: Google describes it as a demonstration project without official product support. Budget for tuning it to your stack. In September 2026 one author rewrote the reference harness on ADK, hardened the sandbox for the reproduction stages and added a campaign planner, still without a tagged release; pin the commit you reviewed.
  • Limits in this sandbox: the reproduction and patching skills expect Docker inside the agent environment, which this guest does not provide. Use the text-only stages and record what you skipped. Mantis also writes working files into the target tree; reset the VM, import the source and reinstall the skills before a comparison run. Local skill runs do not validate its full pipeline.

These tools are candidates for a separate experiment once their setup fits your target. They need separate integration work with the playbook sandbox.

Repository · Apache-2.0

Threat modeling Discovery PR review Reproduction Patching

The CLI and TypeScript SDK behind OpenAI’s hosted Codex Security service, formerly Aardvark. A scan drafts a repository threat model, searches for vulnerabilities against it and validates candidates before reporting them. Separate commands patch selected findings, verify the fix in a read-only sandbox and open a draft pull request. Diff and working-tree scans review changes, and the GitHub Actions example uploads SARIF to code scanning. Choose it when one tool should cover every review stage and OpenAI’s Codex runtime is acceptable.

Setup needs Node.js 22.13 or later, Python 3.10 or later, and either a ChatGPT login or OPENAI_API_KEY. Amazon Bedrock, OpenRouter and Fireworks are documented alternatives for the model, with a limit that decides what you get: the scan delegates to sub-agents (a baseline auditor, an architecture review, focused investigators), and the bundled Codex runtime sends their briefs as encrypted content that only OpenAI’s API unwraps. Through OpenRouter the sub-agents receive empty tasks and report themselves blocked (Codex issues 34833 and 37237, open as of October 2026); the CLI refuses to disable delegation. Setting multi_agent_version: v1 in the Codex overrides made the parent run the whole scan alone in one trial on 2026-10-20, with the same six finding titles as the delegated run on OpenAI’s own model but a finding record the open-weight model could not write in the required shape. Treat other providers as single-agent runs. The runtime also checks the ChatGPT account’s Daybreak access before the first scan action and prints a warning when there is none; the check is scripted and does not change which model runs. The documented default model is gpt-5.6-sol at xhigh effort, with gpt-6.1-sol selectable; the next tagged release switches the default to GPT-6 Sol, so record the resolved model per run. --max-cost bounds an estimate from the CLI’s bundled price table: with a ChatGPT login nothing is billed against it, and for a model outside the table the scan refuses to start with the flag set. Cap spend at the provider instead. Some findings and full-repository scans require Trusted Access for Cyber; see Choosing a model for how that approval is scoped. The package ships about weekly: September 2026 releases added GitLab merge requests, custom severity rubrics, shared CLI and SDK scan settings, Terraform inventories and Bedrock examples for GitHub Actions and Azure Pipelines. The hosted service stays labelled research preview. On September 29, 2026 OpenAI announced Codex Security Cloud with Daybreak Blue models included for Pro, Business, Enterprise and Edu plans; that does not extend to the CLI, which still needs your own approval. OpenAI’s March 2026 announcement reports lower false-positive rates and CVEs found in open-source projects; these are vendor claims, not results validated in this playbook.

Scans run with your operating-system permissions and inherit your environment, so remove unrelated credentials first; the scan reads the whole working copy, ignored files included, and in one trial reported the operator’s own API key from a local .envrc as a finding. Reports and logs are not redacted and can contain source and secrets, and the runtime’s shell snapshots keep the environment of every session. Coverage can be partial; the documentation asks you to read deferred areas before treating a scan as evidence of review. The default validation step is model-driven: one run reviewed sources only, another built a venv and tried to install the target’s dependencies, so allow the target’s package registry if you want that path to succeed. Long single outputs can exceed the bundled runtime’s five-minute stream idle timeout; it retries five times, and a --codex 'model_providers.openai.stream_idle_timeout_ms=...' override lengthens the window. Patching needs the Codex sandbox and refuses to start without it, and --external-sandbox hands isolation to your container. --create-pr commits every verified patch of one invocation to a single branch named after the scan and opens one draft pull request from it; there is no option for one pull request per finding, and a second --create-pr from the same scan fails because the branch exists. Before pushing, it looks up pull requests for that branch name in any state and reports the first match as its result, so a closed pull request on the branch name silently stops publication. For separate pull requests per finding, patch without the flag and open each one yourself.

In this playbook’s sandbox, a codex provider VM with the npm registry ran a full standard scan of the seeded forum on a ChatGPT Plus seat without any allowlist change, finding three of four seeded bugs. The same scan through OpenRouter on GLM-5.3-Flash looped on tool calls, spent more, and was stopped with one finding, while DeepSeek V4.1 Flash completed in nine minutes with two; the acceptance record has the run. Patching inside the guest is untested.

Repository · MIT

Code auditing Alert triage

A framework for repeatable agent workflows, including CodeQL-assisted review and alert triage. A useful candidate if you already have CodeQL databases or scan results. Setup includes taskflows and any required analysis services; the framework’s license does not cover separate CodeQL access requirements.

Repository · MIT framework; component licenses vary

Discovery Repair

An orchestration framework for cyber-reasoning systems, including work from the AIxCC ecosystem. Consider it for sustained fuzzing and repair campaigns, especially on OSS-Fuzz-compatible targets. Choose and configure a CRS as well as the target build; this is a larger integration project than adding a review skill.

Repository · AGPL-3.0

Fuzzing Repair

A system from the AI Cyber Challenge (AIxCC) that combines fuzzing, analysis and patches. Consider it when your project can support a fuzzing campaign and you have capacity to operate its services. It has been tested in competition; setup effort and patch quality still need testing on your code.

Repository · Apache-2.0

Discovery PR review Reproduction Patching Offensive testing for defense

An agentic security tool for source-code review and live application testing. It documents reconnaissance, vulnerability discovery, PoC validation and reporting, plus reviews scoped to PR changes in CI. With source access, its fix workflow covers triage, patching and rescanning findings from the open-source CLI. URL-only testing stops at findings and reports; patching needs the code.

Setup needs Docker, model access and a local codebase or reachable target. Version 1.6 (September 2026) added MCP server connections, a hosted strix cloud service and GLM-5.3 as the default model in its setup examples. Telemetry to third-party services is on by default; set STRIX_TELEMETRY=0 before a run on your code, and keep runs local unless your organisation has approved the cloud service. Start with an isolated demo application. Check PoCs and patches yourself; these are documented capabilities, not results validated in this playbook.

Repository · MIT

Discovery Reproduction Reporting Offensive testing for defense

A multi-agent pentesting system for reconnaissance, vulnerability discovery and exploit attempts. Its example workflow maps endpoints and tests suspected bugs; the documented output includes reproduction steps and vulnerability reports, with Markdown and PDF export. The reviewed documentation does not establish a patch-generation and verification workflow, so patching is not tagged here.

Its service stack includes persistent memory and optional monitoring and knowledge-graph services. Expect more setup than a single review prompt. The last tagged release is v2.1.0 (May 2026). The 2.2.0 code reached the main branch on September 29, 2026 and resolves most of the issues closed in September, but it is not yet tagged. Pin a commit if you install from main. Measure runtime, cost and false positives, and verify reported reproductions before accepting findings.

For Strix and PentAGI, confirm your organisation permits offensive tooling and start with the example application. Isolate the target from production systems and the internet. Check the sandbox’s acceptance requirements before extending its network access; the discovery setup alone does not establish that an offensive workflow is safe.

For validation, a deterministic test remains the default. Every fix needs a regression test that fails before and passes after, plus approval by a human who did not drive the agent. The Validate a finding exercise applies this to one finding from a discovery pass.