The Agent Skills Directory

[DATA_EXFILTRATION]: The skill includes an export-private-key operation (found in bankskills/wallet.py and exposed via the MCP server) that returns the decrypted, plaintext private key of the locally managed Ethereum wallet. While intended as a recovery feature, this capability allows for the exposure of critical credentials to the agent's context, where they could potentially be exfiltrated.
[COMMAND_EXECUTION]: The execution runner in bankskills/runtime/runner.py uses subprocess.run to execute a local shell script (run.sh). This is the standard entry point for the skill but permits the execution of local system commands.
[EXTERNAL_DOWNLOADS]: The skill performs legitimate network operations to well-known services, including the Wise API (api.wise.com) for banking tasks and Ethereum RPC endpoints (e.g., mainnet.base.org) for on-chain transactions.
[PROMPT_INJECTION]: The skill presents a surface for indirect prompt injection due to its processing of external data and high-privilege capabilities.
- Ingestion points: Financial transaction data and profile details from the Wise API (bankskills/core/bank/client.py) and blockchain metadata from RPC calls (bankskills/sweeper.py).
- Boundary markers: None identified in the code that handles or displays external API/RPC responses to the agent.
- Capability inventory: Includes significant capabilities such as send_money (ACH/IBAN transfers), buy_token (Uniswap swaps), and send_token (asset transfers).
- Sanitization: Uses standard address checksumming (Web3.to_checksum_address), but does not perform sanitization on narrative or descriptive data returned from external financial providers.

bank-skill