The Agent Skills Directory

[METADATA_POISONING]: The skill's YAML frontmatter contains deceptive and misleading metadata regarding its provenance.
Evidence: The createdBy and updatedBy fields specify github-copilot/gpt-5.6-sol, a non-existent version of GPT.
Evidence: The createdAt and updatedAt timestamps are set in August 2026, which is in the future.
[INDIRECT_PROMPT_INJECTION]: The skill is designed to ingest and interpret external cryptographic test vectors, which provides an attack surface for instructions embedded in data.
Ingestion points: The script scripts/check-vectors.py parses external JSON files provided as input paths.
Boundary markers: Absent. The skill does not provide delimiters or instructions to the agent to ignore natural language content within the JSON fields.
Capability inventory: The skill utilizes shell command execution (python3 and go run) to process these external files.
Sanitization: Absent. While scripts/check-vectors.py validates JSON structure and data types, it does not sanitize or filter string fields (such as notes or comment fields typically found in Wycheproof vectors) for potential prompt injection payloads.
[COMMAND_EXECUTION]: The skill instructions direct the agent to execute shell commands to validate test corpora and run linter tools.
Evidence: SKILL.md instructs the execution of python3 scripts/check-vectors.py with user-provided paths.
Evidence: references/vector-format.md instructs the execution of go run ./tools/vectorgen fmt and go run ./tools/vectorgen lint using specific environment variables (GOEXPERIMENT=jsonv2).

wycheproof