aig-agent-redteam

Install the Skill into an agent client like Claude Code, CodeBuddy, or Cursor, then ask it to run a security exercise. It models the agent's capabilities and trust boundaries, sends at least 30 injection payloads with adaptive mutation, and reports which boundaries held or failed.

Coverage includes prompt injection, indirect injection via files, webpages, RAG and MCP outputs, tool abuse, data leakage, privilege escalation, SSRF, supply-chain poisoning in Skills and MCP servers, and infrastructure exposure of AI services with fingerprint and CVE matching.

Yes. The code-audit module statically analyzes Skill packages, MCP servers, and plugins, tracing attacker-controlled inputs to privileged sinks like shell, filesystem, and network, and checking tool descriptions and dependencies for instructions that could poison the host agent.

Yes. It supports Claude Code, CodeBuddy, Cursor, and any agent client that supports Skills. Installation uses npx skills add from the GitHub repository, or manual copying into the client's skills directory, requiring only node and git.

Testing requires explicit authorization and scope confirmation before execution. It uses harmless canaries, temporary files, and local mock endpoints instead of real secrets, blocks destructive actions without approval, and keeps all test data on the local machine.

The 30-payload minimum ensures statistical coverage across three sources: at least 10 dataset samples, 10 operator-mutated samples, and 10 hand-crafted first-principles samples. If fewer are sent, the report must state that dynamic coverage is insufficient rather than claiming adequate testing.