The Agent Skills Directory

[EXTERNAL_DOWNLOADS] (MEDIUM): The skill instructs users to install multiple packages (escrowagent-sdk, escrowagent-agent-tools) from an unverified source (cruellacodes). These packages are not hosted within the trusted organizations list and have not been audited for safety.
[CREDENTIALS_UNSAFE] (MEDIUM): The skill documentation requires users to provide AGENT_PRIVATE_KEY and BASE_PRIVATE_KEY via environment variables. While necessary for blockchain operations, providing raw private keys to unverified SDKs and MCP servers poses a significant risk of credential theft if the underlying code is compromised.
[COMMAND_EXECUTION] (MEDIUM): The skill uses npx to execute remote code at runtime (npx escrowagent@latest init and npx skills add cruellacodes/escrowagent). This 'download and execute' pattern from an unverified author allows for arbitrary command execution on the host machine.
[DATA_EXPOSURE] (LOW): The skill references external infrastructure (e.g., escrowagent.onrender.com, escrowagent.vercel.app) for indexing and API services. While not explicitly malicious, interacting with these unverified endpoints with agent-specific data can lead to metadata exposure.
[PROMPT_INJECTION] (LOW): Indirect Prompt Injection risk: The skill processes external data such as 'task descriptions' and 'proof data' within the createEscrow and submitProof methods (SKILL.md).
Ingestion points: createEscrow task object, submitProof data object.
Boundary markers: Absent; no delimiters or 'ignore' instructions for the task metadata.
Capability inventory: Transaction execution, fund locking/releasing, reputation querying.
Sanitization: Absent; the data is passed directly into the tool context.

escrow-agent