The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill reads and summarizes content from experiment files, which could contain adversarial instructions meant to influence the agent's behavior during iterations.
Ingestion points: Reads content from .autoresearch/{domain}/{name}/program.md, .autoresearch/{domain}/{name}/results.tsv, and .autoresearch/{domain}/{name}/config.cfg (SKILL.md).
Boundary markers: No explicit markers or instructions to ignore embedded commands are present.
Capability inventory: Shell command execution (git, cat, python) and potential hand-off to autonomous loops.
Sanitization: No evidence of input validation or content filtering for the files read from the system.
[COMMAND_EXECUTION]: The skill uses placeholders {domain} and {name}, which are populated by user-provided command arguments, directly in shell commands. This represents a potential command injection vector if the platform does not strictly sanitize these inputs.
Evidence: git checkout autoresearch/{domain}/{name} and cat .autoresearch/{domain}/{name}/config.cfg in SKILL.md.

resume