Claude Cybersecurity Code Audit Skill
An AI-powered security audit skill for Claude Code: 8 parallel specialist agents covering OWASP, CWE Top 25, and MITRE ATT&CK with zero configuration.
Strong safety design: explicit prompt-injection defense treating scanned code as data, per-agent tool restrictions (mostly Read/Grep/Glob), and evidence redaction rules for secrets. Deducted for: top-level allowed-tools grants Write/Edit/Bash without a user-confirmation gate or least-privilege statement, recon-phase Bash enumeration, and no rollback/external-effect disclosure.
Methodology is self-consistent with concrete commands, uniform VULN-XXX output, and some designed failure paths (diff fallback, no-IaC returns 100). Deducted for: no test evidence, no reproduced key paths, SKILL.md is truncated mid-Agent-7, and sub-agent behavior/failure feedback on abnormal input is unverifiable.
Clear trigger phrases, arguments (--scope/--focus/--compliance), and supported stacks; purely local execution with no external service dependency, so usable offline. Deducted for: no declared capability boundaries/non-fit ranges and no Chinese-language support statement.
Well-layered docs (SKILL.md + references/), MIT license, SECURITY.md, install/uninstall paths. Deducted for: no versioning/changelog, unclear maintenance ownership, inconsistency between README's '14 languages' and narrative '11', and unverified publisher identity.
Clear goals and professional methodology (context-first, calibrated confidence, false-positive suppression) with plausible marginal value. Deducted for: static review cannot confirm outputs are directly usable; only a video demo link backs representative output, no auditable sample report.
Reference files are auditable primary material and some cited incidents (tj-actions CVE) are public. Deducted for: no committed test suite or CI execution evidence for the skill's key paths, unverifiable statistics ('91% false-positive rate'), and author-authored marketing comparisons vs GHAS.
- This is a static source review only; no scan was executed and the skill's real detection effectiveness is unverified.
- Top-level allowed-tools include Bash/Write/Edit and the orchestrator runs enumeration commands; first use in a controlled environment and review its commands.
- The provided SKILL.md source is truncated at Agent 7; verify completeness of the actually deployed version.
- README suggests a curl | bash install which is a supply-chain risk; prefer cloning and reviewing install.sh.
- No Chinese-language support is declared; no unreachable overseas service dependency exists (fully local), but there is no explicit mainland-China adaptation note.
- No version numbers or changelog; behavior may change across updates without traceability.
What does this skill do, and when should you use it?
This is a Claude Code skill that performs comprehensive code security audits by spawning 8 parallel specialist agents in a single invocation: vulnerability detection, authorization verification, secret scanning, dependency/supply chain analysis, IaC security, threat intelligence (malware/C2/backdoor detection), AI-generated code patterns, and business logic flaws. Findings are aggregated into a structured report with weighted scoring (0-100, grades A-F), per-finding confidence levels, and CWE/OWASP/MITRE mappings. It emphasizes framework-aware false-positive suppression and reasoning about missing security controls, positioning itself as a complement to static tools like GitHub Advanced Security. The skill itself is roughly 23 Markdown reference files (~5,350 lines) with no additional runtime dependencies.
Installed as a 'cybersecurity' skill under skills/ and invoked via /cybersecurity, it works in four phases: 1) GATHER — uses shell commands (find, grep, ls, git diff) to detect the tech stack, languages, frameworks, package managers, IaC and CI/CD setup, enumerate entry points, map trust boundaries, and run STRIDE analysis; 2) ANALYZE — dispatches 8 agents in one parallel message (vulns 20%, auth 15%, secrets 10%, deps 10%, IaC 10%, threat intel 15%, AI code 10%, logic/design 10%), each restricted to Read/Grep/Glob (IaC agent also gets Bash); 3) RECOMMEND — aggregates weighted 0-100 scores, applies an auto-CRITICAL gate, chains attack paths, and optionally maps findings to PCI/HIPAA/SOC 2/GDPR via --compliance; 4) EXECUTE — delivers a report with severity tiers, a remediation priority queue, and MITRE ATT&CK mapping. Supports --scope full|quick|diff and --focus single-dimension deep dives. Built-in rules defend against prompt injection from scanned code and mandate secret redaction in evidence.
- A solo developer or small team running a free full-repository security audit before release (--scope full)
- PR review: scan only changed files before merging (--scope diff) to fold security into daily workflow
- Fast pre-CI checks limited to entry points, auth, secrets, and dependencies (--scope quick)
- Teams with heavy AI-assisted coding who want dedicated detection of missing validation, string-concatenated SQL, and other patterns common in Copilot/ChatGPT output (--focus ai)
- Compliance-driven teams mapping every finding to specific PCI/HIPAA/SOC 2/GDPR requirements (--compliance)
- Single-dimension deep dives when one area is suspect, e.g. supply-chain poisoning investigation (--focus deps or --focus threat)
What are this skill's strengths and limitations?
- Eight parallel audit dimensions, including business-logic flaws and malware/C2 detection that static SAST tools architecturally cannot do
- Framework-aware false-positive suppression across 10 frameworks, with explicit confidence-raise patterns (e.g. dangerouslySetInnerHTML) and confidence-lower patterns (e.g. Django ORM parameterization)
- Every finding carries a confidence tier, CWE/OWASP/MITRE mapping, data-flow path, and concrete before/after fix; the scoring formula is published
- Built-in prompt-injection defense: scanned code is treated as data, and injection attempts themselves are reported as CRITICAL findings
- Zero configuration, MIT licensed, no dependencies beyond Claude Code; full/quick/diff scopes fit different workflows
- Deeply bound to Claude Code-specific mechanisms: allowed-tools, user-invokable frontmatter, and Agent-based subagent spawning do not transfer as-is to other platforms
- Findings are LLM inferences rather than deterministic scans; the repo ships no test suite or benchmark data to back accuracy claims
- Language coverage is inconsistent across sources: the repo description says 11 languages, the README badge says 14, and language-patterns/ actually contains 11 files — real coverage needs verification
- Known-vulnerability checking relies on model memory of notable CVEs (chalk 2025, event-stream 2018, etc.) with no live vulnerability database, so freshness is not guaranteed
- A full 8-agent parallel audit on a large repository will consume significant tokens; cost is not quantified in the source
How do you install this skill?
Three methods per the README: 1) Manual: git clone https://github.com/AgriciDaniel/claude-cybersecurity.git && cd claude-cybersecurity && bash install.sh; 2) Claude Code plugin: claude plugin install cybersecurity; 3) One-liner: curl -fsSL https://raw.githubusercontent.com/AgriciDaniel/claude-cybersecurity/main/install.sh | bash. Uninstall via uninstall.sh or by deleting ~/.claude/skills/cybersecurity. Note: the source does not document where the manual script places files (the uninstall hint shows ~/.claude/skills/cybersecurity); review the script before piping it to bash.
How do you use this skill?
After installing, invoke from a Claude Code session: /cybersecurity for a full audit; --scope quick for a fast scan; --scope diff to review only changed files; --focus threat (or vuln|auth|secrets|deps|iac|ai|logic) for a single-dimension deep dive; --compliance pci to add compliance mapping. Natural-language triggers like "security audit" or "check for vulnerabilities" also work per the SKILL.md description. The only requirement is Claude Code itself — no other configuration needed.
How does this skill compare with similar options?
The README explicitly benchmarks against GitHub Advanced Security: the skill claims to cover what GHAS architecturally misses — business logic flaws, missing authorization enforcement, race conditions, AI-generated code patterns, semantic obfuscated-secret detection, and MITRE ATT&CK-mapped threat intel — for free versus GHAS's $49/committer/month. Keep in mind GHAS is a deterministic, reproducible engine, while this skill is heuristic LLM analysis; the author positions them as complementary rather than competing.