The Agent Skills Directory

[METADATA_POISONING]: The skill is described as a 'vulnerability analyst' tool for testing theft resistance, but its configuration and troubleshooting sections focus heavily on active evasion of security controls.
[PROMPT_INJECTION]: The Troubleshooting section in SKILL.md instructs the agent on how to bypass platform constraints. Specifically, it suggests using 'multiple accounts' to circumvent rate limiting and slowing query rates or varying patterns to avoid triggering detection alerts.
[DATA_EXFILTRATION]: The core functionality involves systematically querying APIs to reconstruct proprietary models. The assets/extraction-config.yaml file defines 'detection_evasion' strategies such as distributed queries and query obfuscation (adding noise tokens, paraphrasing) to facilitate large-scale data harvesting without detection.
[INDIRECT_PROMPT_INJECTION]: The skill possesses a vulnerability surface for processing untrusted data. (1) Ingestion points: External API responses are ingested in SKILL.md (response = target_api(query)) and used as training data. (2) Boundary markers: No delimiters or warnings are used to separate API responses from the agent's internal logic. (3) Capability inventory: The skill performs extensive network operations via the target_api parameter to collect model responses. (4) Sanitization: There is no evidence of sanitization or filtering of the content received from external APIs.

model-extraction