The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes external data from JSONL transcript files which can be provided by users to the verification script.
Ingestion points: The scripts/verify_provably_fair.py script reads and parses content from a file path provided via the --input argument.
Boundary markers: The script uses json.loads() to parse lines, providing structural separation between the data and the script's logic.
Capability inventory: The script performs cryptographic calculations and outputs the results to stdout. The agent is then instructed by SKILL.md to summarize these findings in a sign-off report.
Sanitization: The script performs basic validation of field existence and types (casting to int or str), but it does not sanitize the content of the transcript fields. If a transcript contains malicious text designed to influence the agent, it could manifest when the agent generates the Verification Findings or Residual Risks sections of the output contract.
[COMMAND_EXECUTION]: The skill executes a local script scripts/verify_provably_fair.py which takes sensitive inputs via command-line flags.
Evidence: The SKILL.md file defines a command python3 scripts/verify_provably_fair.py --server-seed "<secret>" .... Passing sensitive cryptographic seeds as command-line arguments is a common practice for local tools, but it can expose the secrets to other users or monitoring processes on the host system via the process table.

rng-crypto-specialist