The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill establishes an attack surface where the agent is instructed to ingest and analyze untrusted data from external server responses.
Ingestion points: Responses from target servers accessed via http_request and processed using analyze_response in SKILL.md (Phase 0).
Capability inventory: The agent is authorized to use http_request for network interaction and python3 for payload generation and executing automation tools.
Boundary markers: The skill lacks instructions for the agent to strictly separate target data from its own internal logic, which could allow a malicious target to influence agent behavior through its output.
Sanitization: No sanitization or validation logic is defined for data received from external sources before the agent processes it or uses it to determine subsequent actions.
[COMMAND_EXECUTION]: The skill instructions and behavior rules in AGENT.md encourage the use of local command-line tools such as python3, tinja, and sstimap. This involves the execution of commands on the agent's host environment to fulfill the penetration testing methodology.

ssti-methodology