The Agent Skills Directory

[PROMPT_INJECTION]: The instructions explicitly direct the agent to circumvent platform security controls intended to protect secrets. By instructing the agent to use os.environ to bypass the '$-expansion' block, the skill attempts to evade safety filters that prevent sensitive data from being exposed in logs or command history.\n- [DYNAMIC_EXECUTION]: The skill uses large blocks of Python code embedded in the markdown, executed via shell HEREDOCs and one-liners. This dynamic generation and execution of code for core logic such as ledger management and pricing is a significant security risk.\n- [INDIRECT_PROMPT_INJECTION]: The skill ingests data from external marketplace APIs (ingestion points include market, price, and earnings endpoints) that influence automated actions. The absence of boundary markers and sanitization logic for these inputs creates an attack surface for indirect prompt injection.\n- [COMMAND_EXECUTION]: The skill requires the agent to perform numerous complex shell operations involving network requests and file system modifications, combined with instructions to bypass platform safety mechanisms.\n- [CREDENTIALS_UNSAFE]: The skill manages multiple high-value credentials for AWS, Google Vertex, and Bankr. The explicit guidance to bypass security features designed to protect these secrets directly increases the risk of credential exposure.

compute-resell