The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes untrusted input from external files (e.g., HTML landing pages, email drafts) and interpolates this content directly into prompts sent to the Anthropic API during its optimization rounds. An attacker could embed malicious instructions within these input files to attempt to override the agent's behavior or exfiltrate the environment's API key.
Ingestion points: The script autoresearch.py reads local file content specified by the --input argument.
Boundary markers: The script utilizes triple dashes (---) as boundary markers in its prompts to separate user content from instructions, which provides some mitigation against accidental instruction following.
Capability inventory: The skill possesses network capabilities (Anthropic API interaction) and file-system write permissions (creating reports and experiment logs in the data/ directory).
Sanitization: No explicit sanitization, filtering, or escaping of the input content is performed before interpolation into the LLM prompts.
[EXTERNAL_DOWNLOADS]: The skill requires the installation of the anthropic Python package, which is the official library for interacting with Anthropic's AI services.
[COMMAND_EXECUTION]: The instructions and README documentation guide the user or agent to execute shell commands for dependency installation and for running the optimization script itself.

autoresearch