akoita/security
Project-agnostic security skills for repository audits, diff review, deterministic scanning, supply chain, threat modeling, smart contracts, and AI systems.
Security review for LLM applications, AI agents, MCP servers and clients, the AI and ML supply chain, and repositories that ship agent skills or plugins. Covers prompt injection and indirect injection, exfiltration paths, agent permission and sandbox hardening, MCP authorization requirements, model, dataset and pickle provenance, AI red-teaming tool selection, and restricted model-assisted vulnerability analysis. Skip it when no model or agent sits in the trust path and a conventional application, infrastructure, or dependency review applies, and skip it for legal or compliance sign-off.
Run a repository-wide security audit and own the shared doctrine — severity, evidence, triage, suppression, reporting — that the other security skills restate. Use for a deep audit of a whole codebase or a large subsystem, for turning scanner output into evidence-backed findings, and for producing a security report. Do not use for a pull-request diff (use security-review), for running the toolchain alone (use security-scan), or for dependency, CI, threat-model, smart-contract, or AI-system questions that have their own skill.
Review a diff or pull request for security-relevant changes and return falsifiable, inline, advisory comments. Use when the unit of work is a change — a branch, a pull request, staged edits — and the question is whether it introduces risk. Do not use as a blocking gate, for a whole-repository audit (use security-audit), for running the deterministic toolchain (use security-scan), or for dependency, threat-model, smart-contract, or AI-system questions that have their own skill.
Run the free deterministic security toolchain — SAST, ecosystem linters, SCA, secrets, IaC, DAST, fuzzing — normalize the output to one digest, and handle each tool's exit codes and suppression syntax. Use when the task is choosing, installing, invoking, or wiring scanners into pre-commit, CI, nightly, or release. Do not use for reasoning about a diff (use security-review), for a judgment-driven repository audit (use security-audit), or for supply-chain, threat-model, smart-contract, or AI-system questions with their own skill.
Audit Solidity and EVM smart contracts, web3 protocols, and on-chain code. Use for smart contract review, Solidity or DeFi audit, proxy and upgradeability checks, invariant and property fuzzing design, oracle and price manipulation review, ERC-4337 or EIP-7702 account abstraction, Permit2 and signature replay, L2 and cross-chain risk, and for running Slither, Aderyn, Echidna, Medusa, Halmos, or Foundry invariants. Skip it for off-chain application, cloud, container, or dependency security, for gas golfing, and for tokenomics or economic-design review with no security question attached.
Harden a project's software supply chain: audit CI/CD workflows, pin actions to commit SHAs, control package-manager and install-script risk, enforce lockfiles and release cooldowns, emit an SBOM, sign artifacts and produce build provenance, and work through the repository checklists and regulatory obligations that follow (OpenSSF Scorecard, OSPS Baseline, EU CRA, NIST SSDF, OWASP ASVS and SAMM). Use when reviewing workflows, dependency intake, release pipelines, or compliance readiness. Do not use for application code vulnerabilities (use security-audit for a repository audit or security-review for a diff), for running the general scanner toolchain (use security-scan), or for LLM, agent, and MCP risk (use security-ai).
Build a repository-grounded threat model: extract the system model from the code, derive trust boundaries, assets and entry points, calibrate attacker capabilities, enumerate abuse paths, rank them by likelihood and impact, and separate existing mitigations from recommended ones. Use for threat modeling, attack surface analysis, trust boundary review, abuse path enumeration, STRIDE or LINDDUN design review, and model-as-code with pytm or threagile, including systems that contain AI agents. Do not use to find concrete vulnerabilities in code (use security-audit for a repository audit or security-review for a diff), to run scanners (use security-scan), or for the detailed LLM, agent, and MCP control checklists (use security-ai).