ReadBeforeRunAgentic Security Field Guide

Catalog / Plugin

Trail of Bits Agentic Actions Auditor

Draft · awaiting review

Claude Code plugin that statically audits GitHub Actions workflows which invoke AI coding agents for injection and permission flaws.

For Agent supply-chain scan, Code review and static analysis (SAST)Works with Claude Code, CodexMaintained trailofbits, last push 2026-10-15

PluginA plugin bundles skills, commands, hooks or MCP servers for one AI client. by trailofbits · read from the code at 5c99c03 (1.3.2) on 2026-10-19The date we last read the tool’s files. The code after “at” names the exact version (the git commit) we read; a newer version may differ. · Not run yet.Whether we have run the tool in our lab. “Did more than it said” means the run used rights its description does not name; “did no more than it said” means it stayed within them.

Code showsWhat a reading of the code found the tool can do, at the checked version, in the words of the rights table in the record. It is a ceiling, not what the tool did. A right marked unknown has no evidence either way, and that never counts as harmless.

What the code shows, worst first. No run advice yet: the one word needs a read under the full rule.

  1. Execution. Bash with no command pattern in allowed-tools permits any shell command.

    SKILL.md L4

  2. Approval. Frontmatter pre-approves Bash with no command pattern, which equals broad pre-approval.

    SKILL.md L4

Before first useSettings for your AI client that limit the tool to the rights its task needs; derived from the code unless marked tested. derived from the code, not tested

  1. Claude Code: scope Bash to gh api and gh auth status only; the skill itself never needs write access.

What we checked, and what we foundDisagree?

Plugin · code showsreach 2 · open
012102FilesNetworkExecutionSecretsPersistenceApproval

What the code can reach, 0 to 3 per right; a dashed point means no evidence, never zero. The three readings

Is it for you?

Made for
  • Detects attack vectors where attacker-controlled input reaches AI agents running in CI/CD pipelines.
  • Produces a structured findings report per attack vector with severity ratings and remediation guidance.
Use it when
Not for
  • Not for runtime prompt injection testing; the author calls it static analysis guidance, not exploitation.
  • Not for non-GitHub CI/CD systems such as Jenkins, GitLab CI or CircleCI.
  • Not for auto-fixing or modifying workflow files; the skill reports findings and does not modify files.
You need
  • Claude Code with the trailofbits/skills plugin marketplace added; the root README also documents Codex
  • The GitHub CLI (gh) for remote analysis mode and remote cross-file fetches
  • An authenticated gh session (gh auth login) for remote fetches

+1 more

Details

Install

linked, not distributed: the Field Guide does not bundle agents or MCP servers

Upstream

/plugin marketplace add trailofbits/skills /plugin menu
Quoted from README.md#L12-18 at the pinned commit; a quotation, not a command to copy.

Run it with least privilege

claude code
scope Bash to gh api and gh auth status only; the skill itself never needs write access
note
derived from the code, not tested in the lab

derived, not tested: proposed from the code analysis

End of the card. Below: what we checked, and what we found.MarkdownJSON

The record · every value with its evidence

What we checked, and what we found.

Profile read at commit 5c99c03 on 2026-10-25: 47 facts, each with its quoted line confirmed by script; not installed, not run in the lab.Each fact was read from the tool's files at the checked version and links to the line it comes from. Where nothing was found, the page says so instead of assuming “none”.

Selected by criteria v0.1, approved by Ilja Schlak: awaiting approval. Each section opens with the answer; the file and line of every fact follow it.

What it can change the capability groups as upstream groups them, worst first

Runs commands, according to the capabilities read.

GroupWhat it coversEffectOn by default
Step 0: Determine Analysis ModeParses owner/repo or a GitHub URL with optional ref and fetches the workflow directory listing and file contents with gh api (remote analysis mode). SKILL.md L68-74reads, runs commandsno, optional
Cross-File ResolutionResolves local composite actions and local or remote reusable workflows referenced by uses:, one level deep, reading from disk or via gh api. cross-file-resolution.md L15reads, runs commandsyes
Step 1: Discover Workflow FilesGlobs .github/workflows/*.yml and *.yaml at the repository root and reads each workflow file. SKILL.md L111readsyes
Step 2: Identify AI Action StepsMatches each step's uses: against five known AI action references and records file, job, step, action reference and type. SKILL.md L124readsyes
Step 3: Capture Security ContextCaptures with: fields per action type, trigger events, env: blocks at three scopes and permissions: blocks for each AI action step. SKILL.md L164readsyes
Step 4: Analyze for Attack VectorsChecks vectors A to I (injection paths, eval of AI output, dangerous sandbox configs, wildcard allowlists) using the per-vector reference files. SKILL.md L232readsyes

Where it sends data and who can connect to it

2 outside connections found in the files read: GitHub API (named repository), GitHub API (referenced repositories).

Listens for no connections in the files read.

Other connections not ruled out: read in the documentation and the cited code, not searched in full.

How it runs generated from the files read, never drawn by hand

Markdown instructions (SKILL.md plus reference files) that the AI client's model follows with Read, Grep, Glob and Bash; it reads local .github/workflows files or calls gh api against the GitHub Contents API for remote repositories and referenced reusable workflows. SKILL.md L4

Your AI client

Parts that live inside the AI client: skill instructions, client configuration.

Your machine

Programs and files on the machine where you run the tool.

Outside your machine

Remote services: vendor APIs, model providers, targets, telemetry endpoints. Data sent here leaves your machine.

The six rights, read three waysThe six rights a tool can take (files, network, execution, secrets, persistence, approval), each read three ways and rated 0 to 3. 0 never means safe; “?” means no evidence. what the author says, what the code shows, what one lab run showed

Level 2 (open), read from the code. Not run in our lab yet.

RightAuthor sayswhat the author saysCode showswhat static analysis findsRun showedwhat one lab run showed
Fileswhere does it read and write?
Networkwhom does it talk to?
Executionwhat does it run?
Secretswhich credentials does it touch?
Persistencewhat survives the run?
Approvaldoes it ask before acting?

0 contained · 1 scoped · 2 open · 3 privileged · ? no evidence · 0 never means safe · zero means: none declared (Declared), none found (Code), none observed (Lab)

Author says: the evidence

  • Files · 0SKILL.md L26 author states the skill reports findings and does not modify files
  • Network · 1SKILL.md L50 author declares a remote mode for a GitHub repository the user names
  • Execution · 1SKILL.md L99-101 author limits Bash to gh api and gh auth status
  • Secrets · 1SKILL.md L91-92 author says remote mode needs the user's own gh login
  • Approval · 2SKILL.md L4 frontmatter pre-approves the Bash tool with no command pattern

Code shows: the evidence

  • Files · 0none found: patterns fs-v0.1 in 5 files at 5c99c03
  • Network · 1cross-file-resolution.md L88 gh api fetches referenced workflows from the GitHub API
  • Execution · 2SKILL.md L4 Bash with no command pattern in allowed-tools permits any shell command
  • Secrets · 1SKILL.md L101 the only credential touch is the user's own gh login
  • Persistence · 0none found: patterns persist-v0.1 in 5 files at 5c99c03
  • Approval · 2SKILL.md L4 frontmatter pre-approves Bash with no command pattern, which equals broad pre-approval

Valid for commit 5c99c03 (1.3.2) · checked 2026-10-19 · method reach-v0.1The published rules we follow: how tools are selected, how their rights are read and rated, how a lab run is done.

Limits and risks what the files say that fits nowhere else, and what this reading does not cover

What you need and what the documented installer does

KindWhat
AI clientClaude Code, Codex
RuntimeClaude Code with the trailofbits/skills plugin marketplace added; the root README also documents Codex. README.md L12
Other toolsThe GitHub CLI (gh) for remote analysis mode and remote cross-file fetches. SKILL.md L68
CredentialsAn authenticated gh session (gh auth login) for remote fetches. SKILL.md L91
CredentialsPrivate repositories are reachable when the gh token in use has permissions for them. SKILL.md L92

What the documented installer does

Lab runs one run is an observation, not a benchmark

No lab run yet. Once we run it, this section shows what it did and whether it did more than it said.

About this entry

Type
Plugin · Agent security · works with Claude Code, Codex
Source
github.com/trailofbits/skills · plugins/agentic-actions-auditor · CC-BY-SA-4.0 · manifest plugin.json
Version readThe install command installs exactly the version we checked (the pinned git commit), so what you get is what the card describes.
5c99c03 · 1.3.2 · last upstream push 2026-10-15 · no release since the pinned commit
In the catalog
listed · added 2026-10-19 · last checked 2026-10-19 · next check 2027-01-17 · fresh
ScannersA scan of the tool's files with a scanner for malicious patterns. “Pending” means it has not run yet.
cisco-skill-scanner: not run, pending
Changes
none since the entry was added on 2026-10-19
MethodThe published rules we follow: how tools are selected, how their rights are read and rated, how a lab run is done.
v0.1 · reach-v0.1 · badge · the rules

Disagree?

Dispute a value with evidence (file, line, commit): write to hello@readbeforerun.ai. A dispute is checked against the pinned commit; a changed value is logged under Changes.

This entry as Markdownas JSON

How this entry was checked, and how to cite it

Selected by the published criteria (method v0.1). What the author says comes only from author statements; what the code shows comes from the code and manifests at commit 5c99c03, each with one evidence line. A “?” means no evidence, never zero; zero never means safe.

The descriptive profile (what it does, can do, how it runs, what it needs) was read through the same isolated pipeline at commit 5c99c03 on 2026-10-25: one reader proposed every fact with a quoted line, a script confirmed each quoted line (47 of 47), and a second reader checked each sentence against its lines. Where nothing was found, the page says so instead of assuming “none”.

Known issues and limits: remote server-side behaviour stays “?”; the lab is not your production; static analysis can miss runtime behaviour.

Placement, badges or fast tracks are not for sale. The author is employed by an IT consultancy; the company has no influence on selection, execution or result.

Cite this:

Ilja Schlak. "Trail of Bits Agentic Actions Auditor". Agentic Security Field Guide, checked 2026-10-19, method reach-v0.1. https://readbeforerun.ai/catalog/trailofbits-agentic-actions-auditor/

The method · the rules

For maintainers

Badge for your README (reach 2 · not run yet):

[![reach 2 · not run yet](https://readbeforerun.ai/badge/trailofbits-agentic-actions-auditor.svg)](https://readbeforerun.ai/catalog/trailofbits-agentic-actions-auditor/)

Dispute a value with evidence (file, line, commit): write to hello@readbeforerun.ai. A dispute is checked against the pinned commit; a changed value is logged under Changes.

Claim the entry with a fieldguide.yaml in your repository: planned.

Other tools for the same task

Tasks: Agent supply-chain scan, Code review and static analysis (SAST). The task pages compare every tool side by side.
  1. 12232?
    Cisco MCP Scanner PluginPlugin · by cisco-ai-defense · reach 3, privilegedDraft · awaiting review

    Claude Code plugin that scans MCP servers and tools for prompt injection, tool poisoning and other risks via the Cisco MCP scanner.

    Can read credential stores · talk to any host · run shell commands or eval · +1 more · Approval: no evidence · ran once, 2026-11-07: did more than it said · checked 2026-11-07

    Code shows
  2. 202210
    Cloudflare Security Audit SkillSkill · by cloudflare · reach 2, openDraft · awaiting review

    Agent skill that runs a multi-phase, source-first security audit of a codebase and verifies each candidate finding.

    Can write outside the project · run shell commands or eval · read broad tokens or the environment · not run yet · checked 2026-10-19

    Code shows
  3. 222?22
    Trail of Bits Static AnalysisPlugin · by trailofbits · reach ≥2, openDraft · awaiting review

    Claude Code plugin that drives CodeQL, Semgrep and SARIF parsing to find security vulnerabilities in a codebase.

    Can write outside the project · talk to any host · run shell commands or eval · +2 more · Secrets: no evidence · not run yet · checked 2026-10-19

    Code shows