The Agent Skills Directory

[COMMAND_EXECUTION]: The skill is designed to autonomously execute shell commands defined in a configuration file (program.md) as part of an experimental loop.
Evidence: SKILL.md and references/implementation-guide.md describe running a 'target system' and 'predictor' using invocation commands provided by the user. The agent executes these commands to sample outputs across a parameter space.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection because it ingests and processes untrusted data from external command outputs and web search results within its autonomous loop.
Ingestion points: Data enters the agent context through program.md, target/predictor command outputs, and web results from academic and developer platforms.
Boundary markers: There are no specified delimiters or instructions to treat external data as untrusted or to ignore instructions embedded within it.
Capability inventory: The skill has extensive capabilities including shell command execution, file system access (writing logs and reports), and network access via web search tools.
Sanitization: No evidence of input validation or content sanitization is present for data retrieved from external sources before it is processed by the agent.
[EXTERNAL_DOWNLOADS]: The skill performs automated literature searches using well-known services to verify findings.
Evidence: references/discovery-mode.md details a literature check step involving searches on platforms like Google Scholar, arXiv, GitHub, and Wikipedia. These operations target well-known services and are documented as part of the primary discovery workflow.

anomaly-hunter