The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill contains hardcoded cryptographic keys used for HMAC-based hashing operations. In scripts/process.py, the BloomFilterEncoder class uses b"default_key" as a default value for its secret_key parameter. Additionally, the execution block in the same file uses a hardcoded key = b"shared_secret_key_2026". Hardcoding keys used to protect personally identifiable information (PII) allows attackers with access to the code to perform dictionary attacks or reverse-engineer the hashed values.
[INDIRECT_PROMPT_INJECTION]: The skill handles untrusted data inputs representing record attributes (such as names, dates of birth, and addresses) from external datasets for the purpose of entity resolution. \n- Ingestion points: The encode_record and hash_dataset methods in SKILL.md and scripts/process.py accept dictionaries of attribute strings from external parties. \n- Boundary markers: The skill does not implement boundary markers or instructions to ignore embedded prompts within the record data. \n- Capability inventory: The skill performs string normalization, n-gram generation, and cryptographic hashing. It does not possess network transmission or filesystem write capabilities in the provided code. \n- Sanitization: Inputs are stripped of whitespace and converted to lowercase, which provides basic normalization but does not mitigate prompt injection risks if the resulting strings were to be processed by an LLM downstream.

privacy-record-linkage