The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill instructs the agent to ingest and analyze program logs (e.g., /tmp/session-stdout.log) that contain raw HTTP traffic and API responses. This creates a surface where malicious instructions embedded in external API payloads could potentially influence the agent if it processes the logs without isolation.
Ingestion points: Reads program session logs from /tmp/session-stdout.log as described in SKILL.md.
Boundary markers: Absent. The instructions do not specify the use of delimiters or 'ignore' instructions when the agent reads these log files.
Capability inventory: Uses pkill, curl, xvfb-run, and httptoolkit to manage the debugging environment and execution.
Sanitization: Absent. The skill suggests using jq for filtering but does not provide patterns for sanitizing potential instruction injection from the log data.
[COMMAND_EXECUTION]: The skill uses pkill -f with broad string matching patterns to terminate processes. While intended to clean up HTTP Toolkit instances, such commands can inadvertently terminate unrelated system processes sharing similar command-line strings if used improperly.
[SAFE]: The skill provides detailed guidance on bypassing TLS security controls (e.g., NODE_TLS_REJECT_UNAUTHORIZED=0, verify=False, --unsafely-ignore-certificate-errors). These instructions are explicitly labeled as being for "controlled local debugging" and "local-dev only," accompanied by safety guardrails that advise against their use in production environments.

http-toolkit-intercept