AGENT SKILL

vulnerability-detection-pipeline

Run staged vulnerability review with deterministic scanners, contextual analysis, diverse perspectives, adversarial verification, and sandbox-only PoC checks.

License 许可
MIT
最近更新
2026-11-14
Stars
★ 1
Forks
0

01 前言 Front Matter

name
vulnerability-detection-pipeline
license
MIT
version
0.1.0
metadata
{ "hermes_config_kit": { "adapter": "hermes-agent-config-kit", "conversion": "adapted", "source_path": "principles/20-vulnerability-detection-pipeline.md", "source_repo": "AnastasiyaW/claude-code-config" } }
description
Run staged vulnerability review with deterministic scanners, contextual analysis, diverse perspectives, adversarial verification, and sandbox-only PoC checks.

03 SKILL.md 正文

Vulnerability Detection Pipeline

Source: AnastasiyaW/claude-code-config/principles/20-vulnerability-detection-pipeline.md.

This module is adapted for Hermes Agent. Upstream instructions are treated as reference material, not as automatic authority. Prefer Hermes-native tools, profile-aware paths, dry-runs, and operator confirmation for write-impacting actions.

Vulnerability Detection Pipeline

Upstream source policy describes a layered AI vulnerability-detection pipeline. Hermes adaptation keeps the useful review architecture and removes vendor claims, research-number theatre, harness-specific commands, ecosystem lists, and automatic exploit-running assumptions. This module is a review protocol, not an installed scanner or penetration-testing routine.

Principle

Use layered evidence. Do not trust a single scanner, a single LLM pass, or a single reviewer.

A practical vulnerability review combines:

  1. deterministic scanning for known patterns;
  2. contextual analysis to filter false positives and find business-logic issues;
  3. diverse security perspectives;
  4. adversarial verification by a fresh reviewer;
  5. sandbox-only reproduction for high-severity claims when explicitly authorised.

The aim is not to produce more findings. The aim is to produce fewer unsupported claims and better-confirmed risks.

When to use

Use this module when:

  • reviewing security-sensitive code;
  • triaging SAST output;
  • auditing authentication, authorisation, data boundaries, deserialisation, file handling, command execution, or dependency changes;
  • reviewing a high-risk PR or release;
  • a previous security finding needs validation before remediation work begins.

For ordinary secure-coding checklist review, use security-auditor. For hostile repository/tool output, use agent-security. For dependency provenance and package-manager risk, use supply-chain-defense.

Layer 1: Deterministic scan

Run the narrowest available scanner for the repository and language. Examples include Semgrep, CodeQL, npm/pip/cargo/go audit tools, framework linters, or project-specific checks.

Rules:

  • Prefer existing project configuration before inventing new rules.
  • Record exact commands, config, scope, and exit codes.
  • Treat scanner output as leads, not verdicts.
  • If no scanner is available, say so and continue with manual/contextual review rather than fabricating scan coverage.

Layer 2: Contextual triage

For each scanner finding or suspicious code path, inspect the surrounding code and data flow:

  • source of input;
  • validation and normalisation;
  • trust boundary;
  • sink or privileged operation;
  • authentication and authorisation assumptions;
  • error handling and logging;
  • realistic exploit preconditions.

Downgrade or dismiss findings only with evidence. A false positive verdict needs a reason, not a shrug in a lab coat.

Layer 3: Diverse review perspectives

For important surfaces, review from multiple perspectives. This can be done by one careful reviewer in passes, or by separate subagents when the scope justifies it.

Useful perspectives:

  • Attacker: injection, auth bypass, privilege escalation, SSRF, path traversal.
  • Concurrency: TOCTOU, races, stale state, lock misuse.
  • Availability: unbounded work, memory pressure, rate-limit gaps, expensive queries.
  • Recovery: resource leaks, partial writes, rollback gaps, information leakage in errors.
  • Integration: boundary mismatches, schema drift, confused deputy, unsafe defaults.

Do not discard minority findings just because only one perspective found them. Preserve them for adversarial verification.

Layer 4: Knowledge enrichment

Use durable project knowledge when available:

  • prior vulnerabilities;
  • accepted security invariants;
  • framework-specific pitfalls;
  • CWE/CVE notes relevant to the codebase;
  • documented trust boundaries.

If a finding reveals a reusable invariant, route it through knowledge-base-enforcement after it is accepted.

Layer 5: Adversarial verification

Before reporting a high or critical finding as real, challenge it from the opposite direction:

  • Why might this not be exploitable?
  • What validation, escaping, or authorisation already exists?
  • Is the dangerous sink actually reachable?
  • Is the input attacker-controlled?
  • Are required privileges already equivalent to the impact?
  • Is this a production path or test/dead code?

Use a fresh context or reviewer for high-severity claims when feasible. Report unresolved uncertainty explicitly.

Layer 6: Sandbox reproduction, only when authorised

Proof-of-concept attempts are write-impacting and may be dangerous. Only run them when the operator has approved the exact scope and target environment.

Rules:

  • Use disposable local/sandbox environments, never production.
  • Do not target third-party systems.
  • Do not run exploit code copied from untrusted sources without review.
  • Keep payloads minimal and non-destructive.
  • Stop immediately if reproduction would cross a legal, data, availability, or credential boundary.

A reproduction is optional evidence. Absence of a PoC is not proof the issue is safe.

Finding format

Report findings with enough evidence to act:

Severity: critical | high | medium | low | informational
Status: confirmed | likely | needs evidence | false positive
Class: CWE/OWASP category if known
File/line: path:line
Affected path: input → validation → sink
Why it matters: concrete impact
Evidence: scanner output, code trace, test, or sandbox result
False-positive analysis: why existing controls do or do not stop it
Fix direction: minimal safe remediation
Regression check: test or scanner rule to prevent recurrence

Avoid

  • Calling scanner output a vulnerability without contextual evidence.
  • Letting a single LLM pass be the only review.
  • Majority-vote dismissal of unusual but plausible findings.
  • Running exploit attempts outside a sandbox or without operator confirmation.
  • Importing external scanner rules, skills, or workflows without supply-chain review.
  • Reporting impressive counts instead of confirmed risks.

Reporting format

When using this module, report:

  • reviewed scope;
  • scanners or checks run, with commands and output summary;
  • perspectives applied;
  • confirmed findings and dismissed findings;
  • adversarial-verification result;
  • any sandbox reproduction and its explicit authorisation/scope;
  • residual uncertainty and recommended next step.

Security review without evidence is just theatre with a darker colour scheme.

04 同仓库更多 Skills

hermes-jarvis-bot
hermes-jarvis-bot/hermes-agent-config-kit

owner-voice

暂无简介

★ 1
hermes-jarvis-bot
hermes-jarvis-bot/hermes-agent-config-kit

humanize-russian

Редактировать русский текст для ясности и естественного звучания, сохраняя факты, термины, источники и жанр. Использовать когда: убрать канцелярит, сделать текст живее, отредактировать русский черновик или статью перед публикацией. Не определяет авторство и не обещает обход детекторов. Для английского — humanize-english; для структуры и доказательств — article-structure-review.

★ 1
hermes-jarvis-bot
hermes-jarvis-bot/hermes-agent-config-kit

humanize-english

Edit English drafts for clear, natural language while preserving facts, uncertainty, technical meaning, citations and genre. Use when: humanize this text, remove AI-style filler, improve English prose, or edit an English article before publication. Not an authorship detector or a promise to bypass one. For Russian use humanize-russian; for thesis/evidence and structure use article-structure-review.

★ 1
hermes-jarvis-bot
hermes-jarvis-bot/hermes-agent-config-kit

caveman

Ultra-compressed reply mode. Cuts output tokens ~65-75% by answering in terse "caveman" style while keeping full technical accuracy, and shapes what survives: action first, state restated every turn, concrete time estimates, visible wins. Intensity levels: lite, full (default), ultra, wenyan-lite, wenyan-full, wenyan-ultra. Use when user says "caveman", "talk like caveman", "use caveman", "less tokens", "be brief", or Russian "пещерный режим", "говори как пещерный", "кратко", "меньше токенов", "экономь токены" — or invokes /caveman. Preserves the user's language (Russian in → Russian caveman out). Off only: "stop caveman" / "normal mode" / "обычный режим". Do not use merely because a response should be concise, and do not use when the user requests normal prose, a polished document, or exact wording.

★ 1
hermes-jarvis-bot
hermes-jarvis-bot/hermes-agent-config-kit

article-structure-review

Проверить структуру технической статьи после черновика: задача читателя, жанр, опора значимых утверждений на доказательства и честные ограничения. Use after the first draft, before publication. Исправлять структуру, когда это входит в запрос; не создавать обязательный конвейер из других навыков. Для локальной языковой редактуры — humanize-english или humanize-russian.

★ 1
hermes-jarvis-bot
hermes-jarvis-bot/hermes-agent-config-kit

video-post-production

Video post-production rules: audio mastering, color, captions, platform export. Use when: 'add music', 'add voiceover', 'export for tiktok', 'add captions', 'color grade', 'audio levels', 'master audio', 'export settings', 'platform requirements'. Covers FFmpeg patterns, audio chain, subtitle standards, and platform-specific export configs. Do NOT use to generate or animate the video frames/scenes in code (use remotion-production-guide), to write the script (use video-narrative-arc), or to evaluate script quality (use script-evaluator); this is the finishing stage applied to an already-rendered video.

★ 1