Prompt Injection Tester
A red-team harness for agents. It runs a curated library of injection and jailbreak payloads against your agent and reports what broke.
Attack classes
- Direct injection — "ignore previous instructions" and friends
- Indirect injection — payloads hidden in tool output / fetched pages
- Context exfiltration — attempts to leak the system prompt or secrets
- Tool misuse — coaxing unintended
runCommands/networkcalls
Install
bash
npx nuclexa install prompt-injection-tester --target claude-codeRequestsnetworkandrunCommandsto exercise your agent in a sandbox. Run it against your own agents only.
Report
Each payload is scored: blocked, partial, or compromised, with the transcript and a hardening suggestion.