MCP HubMCP Hub
SKILL·D1F8BC

vulnerability-triage-brocards

trailofbits
更新日 1 month ago
6 閲覧
6,918
596
6,918
GitHubで表示
メタgeneral

について

このスキルは、脆弱性レポートを体系的にトリアージし、七つの反証可能な経験則を適用して、受諾、却下、または追加情報の要求を判断します。CVE、バグバウンティ提出物、セキュリティパイプラインからの未加工の発見を評価し、人間によるレビューの前にフィルタリングするために設計されています。脆弱性レポートがより深い調査に値するか、優先順位を付けるべきかを迅速に判断する必要がある場合にご利用ください。

クイックインストール

Claude Code

推奨
メイン
npx skills add trailofbits/skills -a claude-code
プラグインコマンド代替
/plugin add https://github.com/trailofbits/skills
Git クローン代替
git clone https://github.com/trailofbits/skills.git ~/.claude/skills/vulnerability-triage-brocards

このコマンドをClaude Codeにコピー&ペーストしてスキルをインストールします

ドキュメント

Vulnerability Triage Brocards

Systematically evaluate incoming vulnerability reports against 7 principled criteria before committing resources to deeper analysis. Each brocard is a falsifiable test: if a report fails any brocard, document the reason and dismiss or request clarification. If a report survives all 7, escalate it.

The 7 brocards are adapted from William Woodruff's "Brocards for vulnerability triage" (2026).

When to Use

  • Filtering findings from agentic vulnerability discovery pipelines before human review -- the primary use case; most automated runs produce findings that fail one or more brocards and can be dismissed without auditor time
  • Triaging findings during a ToB audit to decide which warrant escalation to PoC development
  • Evaluating third-party CVEs or advisories against a codebase under active audit to decide if they affect engagement scope
  • Reviewing bug bounty submissions or external vulnerability reports for ToB open-source projects
  • Providing structured, defensible justification when recommending a client dismiss or deprioritize a reported CVE

When NOT to Use

  • Hunting for new bugs during an audit -- use other skills
  • Proving exploitability of a confirmed finding -- use a dedicated PoC/exploitability skill
  • Triaging fuzzer crashes in C/C++ -- use a dedicated crash triage skill

Pipeline Position

This skill is the quality gate between automated discovery and human review. Findings that survive triage proceed to PoC development and formal writeup.

flowchart TD
    A([agentic vulnerability discovery]) -->|raw findings| B[vulnerability-triage-brocards]
    B -->|DISMISS| C([Document brocard # and reasoning])
    B -->|NEEDS-MORE-INFO| D([Request specific evidence])
    B -->|ACCEPT| E[PoC / exploitability proof]
    E --> F[vulnerability report writeup]

Triage Workflow

For each incoming vulnerability report, evaluate it against all 7 brocards sequentially. By default, stop at the first DISMISS verdict and report it. If the user requests a full evaluation, continue through all 7 brocards regardless of intermediate failures. For each brocard, record one of three verdicts:

  • PASS -- the report survives this test
  • DISMISS -- the report fails this test; document the reason
  • NEEDS-MORE-INFO -- insufficient evidence to evaluate; specify what is missing

Brocard 1: No Vulnerability Without a Threat Model

Dismiss any report that lacks a coherent threat model. The report must articulate: (a) who the attacker is, (b) what capability the attacker has, (c) how the attacker exploits the behavior, and (d) what harm results.

Reports that describe a code behavior without connecting it to attacker- reachable harm fail this brocard.

Quick test: Can the report answer "an attacker with [capability] can [action] to achieve [impact]"? If not, dismiss or request clarification.

Brocard 2: No Exploit from the Heavens

Dismiss any report where the attacker capabilities required to trigger the vulnerability equal or exceed the impact of the vulnerability itself. If the attacker must already possess the power the exploit would grant, the vulnerability is redundant.

Quick test: Does triggering the exploit require capabilities that already subsume its impact? If yes, dismiss.

Brocard 3: No Vulnerability Outside of Usage

Dismiss any report describing behavior that is theoretically possible but does not occur in actual software usage. Check whether the vulnerable code path is reachable in practice.

Quick test: Is the vulnerable code path exercised by any real caller? If not, dismiss. If the report targets a library, ask if we should check downstream usage.

Brocard 4: No Vulnerability from Standard Behavior

Dismiss any report where the behavior results from correct implementation of a specification. The vulnerability, if any, exists in the standard -- not the implementation.

Nuance: If an implementation voluntarily adopts a stricter posture than the standard requires, and that strictness fails, the implementation is vulnerable even though the standard permits the behavior.

Quick test: Does the specification require or permit this behavior? If yes, the report targets the standard, not the code.

Brocard 5: No Vulnerability from Documented Behavior

Dismiss any report describing behavior that is explicitly documented, especially when the documentation includes security implications or usage caveats.

Nuance: Downstream usage that violates documented guidelines may constitute a valid vulnerability in the downstream project, not the documented component.

Quick test: Does the project's documentation describe this behavior and warn against misuse? If yes, dismiss the report against the project itself.

Brocard 6: No Cure Worse Than the Disease

Dismiss any report whose remediation would cause more harm than the vulnerability itself. Evaluate: (a) severity of the vulnerability in practice, (b) cost and disruption of the proposed fix, (c) blast radius of the remediation (dependency graph, ecosystem impact).

Quick test: Would fixing this cause more disruption than the vulnerability itself? If yes, dismiss or downgrade severity.

Brocard 7: The Report Is Neither Necessary nor Sufficient

A CVE identifier or formal report does not prove a vulnerability exists. Conversely, absence of a report does not prove safety. Evaluate the technical merits independently of report metadata.

Quick test: Strip the CVE number and CVSS score. Does the technical description alone justify action? Judge on evidence, not authority.

Output Format

After evaluating all 7 brocards, produce a structured triage summary:

## Triage Summary: [Report ID or Title]

| # | Brocard | Verdict | Rationale |
|---|---------|---------|-----------|
| 1 | Threat Model | PASS/DISMISS/NEEDS-MORE-INFO | ... |
| 2 | Exploit from the Heavens | PASS/DISMISS/NEEDS-MORE-INFO | ... |
| 3 | Outside of Usage | PASS/DISMISS/NEEDS-MORE-INFO | ... |
| 4 | Standard Behavior | PASS/DISMISS/NEEDS-MORE-INFO | ... |
| 5 | Documented Behavior | PASS/DISMISS/NEEDS-MORE-INFO | ... |
| 6 | Cure Worse Than Disease | PASS/DISMISS/NEEDS-MORE-INFO | ... |
| 7 | Report Sufficiency | PASS/DISMISS/NEEDS-MORE-INFO | ... |

**Overall Verdict:** ACCEPT / DISMISS / NEEDS-MORE-INFO
**Reasoning:** [1-3 sentence justification]
**Next Step:** [escalate to PoC development / request info / close]

Rationalizations to Reject

Guard against these reasoning failures in both directions:

Wrongly Dismissing Valid Findings

  • "It's only reachable in debug mode" -- verify debug mode is truly never enabled in production; many clients ship with debug flags on
  • "The attacker would need local access" -- local access is a realistic threat model for many deployments, especially containerized services
  • "Nobody uses that API" -- confirm with actual usage data, not assumptions; check client's integration tests and deployment configs
  • "The spec allows it" -- check whether the implementation claims stricter behavior than the spec requires

Wrongly Accepting Invalid Findings

  • "It has a CVE, so it must be real" -- Brocard 7 exists for this reason
  • "The CVSS score is high" -- CVSS is a formula, not a verdict
  • "Better safe than sorry" -- Brocard 6 requires evaluating fix cost
  • "We can't prove it's NOT exploitable" -- the burden of proof is on the reporter to demonstrate a threat model (Brocard 1)
  • "Other projects patched it" -- other projects may have different usage patterns (Brocard 3)
  • "We should include it to pad the report" -- ToB reports reflect technical reality, not finding count targets; a dismissed report with documented reasoning is more valuable than a false positive in a final deliverable

Detailed References

For expanded explanations, examples, and edge cases for each brocard, consult references/brocards-detail.md.

GitHub リポジトリ

trailofbits/skills
パス: plugins/vulnerability-triage-brocards/skills/vulnerability-triage-brocards
0
agent-skills
FAQ

よくある質問

vulnerability-triage-brocards Skillとは何ですか?

vulnerability-triage-brocards はtrailofbits が作成した Claude Skillです。Skillは、Claudeが必要に応じて読み込む指示とリソースをまとめ、追加の指示なしで vulnerability-triage-brocards に関連するタスクを実行できるようにします。

vulnerability-triage-brocards をインストールするには?

このページのインストールコマンドを使用してください。vulnerability-triage-brocards をプラグインとして Claude Code に追加するか、リポジトリを skills ディレクトリにクローンし、Claudeを再起動してSkillを読み込みます。

vulnerability-triage-brocards はどのカテゴリに属しますか?

vulnerability-triage-brocards は メタ カテゴリに属します。

vulnerability-triage-brocards は無料で利用できますか?

はい。vulnerability-triage-brocards は AIMCP に掲載されており、無料でインストールできます。

関連スキル

content-collections
メタ

このスキルは、Content Collections(Markdown/MDXファイルを型安全なデータコレクションに変換するTypeScriptファーストのツール)の本番環境でテストされた設定を提供します。Zodバリデーションによる型安全性を実現し、ブログ、ドキュメントサイト、コンテンツ重視のVite + Reactアプリケーション構築時にご利用ください。Viteプラグインの設定、MDXコンパイルから、デプロイ最適化、スキーマバリデーションまで、すべてを網羅しています。

スキルを見る
polymarket
メタ

このスキルは、開発者がPolymarket予測市場プラットフォームを活用したアプリケーション構築を可能にします。API統合による取引や市場データの取得に加え、WebSocketを介したリアルタイムデータストリーミングにより、ライブ取引や市場活動を監視できます。取引戦略の実装や、ライブ市場更新を処理するツールの作成にご利用ください。

スキルを見る
creating-opencode-plugins
メタ

このスキルは、開発者がコマンド、ファイル、LSP操作など25種類以上のイベントタイプにフックするOpenCodeプラグインを作成することを支援します。JavaScript/TypeScriptモジュール向けに、プラグイン構造、イベントAPI仕様、および実装パターンを提供します。カスタムイベント駆動ロジックでOpenCode AIアシスタントのライフサイクルをインターセプト、監視、または拡張する必要がある場合にご利用ください。

スキルを見る
sglang
メタ

SGLangは、高性能なLLMサービングフレームワークであり、RadixAttentionプレフィックスキャッシュを活用したJSON、正規表現、エージェントワークフロー向けの高速で構造化された生成を特長とします。特にプレフィックスが繰り返されるタスクにおいて、大幅に高速な推論を実現し、複雑な構造化出力やマルチターン対話に最適です。制約付きデコードが必要な場合や、広範なプレフィックス共有を伴うアプリケーションを構築する場合は、vLLMなどの代替案ではなくSGLangを選択してください。

スキルを見る