offensive-ai-security
AI and LLM application security assessment aligned with OWASP Top 10 for LLM Applications covering direct and indirect prompt injection, agent tool-call authorization gates, RAG document ACL enforcement, structured output validation, and system prompt isolation.
How do I install this agent skill?
npx skills add https://github.com/smartworldarafath/rea_skill --skill offensive-ai-securityIs this agent skill safe to install?
- Gen Agent Trust Hubpass
This skill provides a structured framework for auditing the security of AI and LLM applications based on the OWASP Top 10. While the instructions define a workflow for analyzing external code and binary artifacts which creates an attack surface for indirect instructions, the skill does not contain any malicious patterns, obfuscation, or unauthorized exfiltration attempts.
- Socketwarn
1 alert: gptAnomaly
- Snykpass
Risk: LOW · No issues
What does this agent skill do?
AI, LLM & Agent Security Audit (OWASP LLM Top 10) (offensive-ai-security)
Authorization & Scope Notice: This skill is part of the REA_Skill + claude-red Unified Security & Reverse Engineering Suite. It is designed strictly for authorized security assessments, defensive architecture reviews, vulnerability triage, bug bounty verification, and remediation engineering on software, infrastructure, and assets you own or have explicit permission to test.
Overview
AI and LLM application security assessment aligned with OWASP Top 10 for LLM Applications covering direct and indirect prompt injection, agent tool-call authorization gates, RAG document ACL enforcement, structured output validation, and system prompt isolation.
Operational Workflow
- Surface & Architecture Inventory
- Identify all entrypoints, trust boundaries, parsers, and configuration files relevant to AI, LLM & Agent Security Audit (OWASP LLM Top 10) (
aidomain). - Combine static code/binary inspection (via REA MCP tools,
rg, AST analysis, or decompilation) with runtime configuration verification.
- Identify all entrypoints, trust boundaries, parsers, and configuration files relevant to AI, LLM & Agent Security Audit (OWASP LLM Top 10) (
- Hypothesis & Vulnerability Verification
- Test each security invariant systematically against known CWE and OWASP patterns.
- Record concrete evidence: affected file/procedure, line numbers or binary offsets, request/response trace, and root cause.
- Impact & Reachability Triage
- Evaluate whether the issue is reachable across trust boundaries, whether defense-in-depth controls mitigate it, and compute an accurate CVSS v3.1 / v4.0 score.
- Remediation & Regression Verification
- Implement or provide production-ready, copy-pasteable remediation code.
- Write an automated regression test or verification command to prove the flaw is resolved with 0 regressions.
Audit & Verification Checklist
| Phase | Check | Verification Method | Expected Secure State |
|---|---|---|---|
| 1. Discovery | Enumerate all ai surfaces and dependencies | Source search / REA binary & bundle inspection | Complete inventory of trust boundaries |
| 2. Input & State | Validate parser behavior, type safety, and state transitions | Boundary testing & static taint analysis | Strict schema validation and invariant checks |
| 3. Auth & Isolation | Verify caller identity, tenant isolation, and least privilege | Horizontal/vertical privilege matrix check | Explicit deny-by-default enforcement |
| 4. Defense-in-Depth | Inspect secondary controls, logging, and compiler/runtime flags | Config & header/binary flag audit | Full hardening and telemetry enabled |
| 5. Remediation | Apply fix and execute automated regression test | npm test / toolchain build & test suite | 0 vulnerabilities, 0 build errors |
Integration with REA MCP & Tooling
- Static & Binary Analysis: Pair with
rea-universal-orchestrator,rea-binary-disassembly,rea-web-bundle-recovery, orrea-mobile-app-recoverywhen source code is unavailable. - Interactive Scoping: Invoke
/rea-grill-mebefore large-scale assessments to clarify scope, target environments, and acceptance criteria. - Reporting: Pair with
offensive-reportingto format findings with CVSS vectors, CWE mappings, and remediation patches.
How can the creator link this skill?
Add the canonical catalog link to the repository README so users can inspect current installs and available audits. The publishing guide covers the complete discovery path.
<a href="https://skillzs.dev/skills/smartworldarafath/rea_skill/offensive-ai-security">View offensive-ai-security on skillZs</a>