The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The skill instructs the user to download and run scripts from unverified sources, specifically jwt-tool from https://github.com/ticarpi/jwt_tool and the Docker image ttervoort/jws2pubkey.
[COMMAND_EXECUTION]: Extensive use of shell commands such as python3, openssl, curl, and docker to perform security testing tasks, frequently incorporating external inputs as arguments.
[DYNAMIC_EXECUTION]: Generates and executes Python code at runtime for token manipulation and JWKS creation, which can be difficult to audit and presents a risk if inputs are compromised.
[DATA_EXPOSURE_AND_EXFILTRATION]: Recommends saving sensitive tokens to local files and demonstrates techniques like jku spoofing that cause the target server to perform outbound network requests to attacker-controlled infrastructure.
[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted JWT tokens provided by the target application and uses them as inputs for command-line tools without explicit sanitization.
Ingestion points: JWT tokens entering agent context in Step 1 Assessment.
Boundary markers: Absent.
Capability inventory: Shell execution (jwt_tool), file writing (engagement/evidence/), and network operations (curl).
Sanitization: Absent.

jwt-attacks