The Agent Skills Directory

SAFE (SAFE): No malicious patterns or security threats were detected. The script implements a keyword-based heuristic to organize paper titles into a taxonomy structure.
Indirect Prompt Injection (LOW): The skill processes untrusted paper metadata (titles and abstracts) which could theoretically contain malicious instructions.
Ingestion points: papers/papers_dedup.jsonl via the read_jsonl function.
Boundary markers: None implemented for the input data or output YAML.
Capability inventory: File system write access via dump_yaml to outline/taxonomy.yml.
Sanitization: The _pretty function performs basic formatting (capitalization and character replacement), but does not specifically sanitize for prompt injection. Given that the script uses algorithmic tokenization rather than an LLM to process this data, the risk of immediate exploitation is negligible.

survey-seed-harvest