The Agent Skills Directory

[PROMPT_INJECTION] (HIGH): Indirect Prompt Injection Surface. Ingestion points: The aptos_intent tool in SKILL.md is designed to interpret natural language instructions, which may include malicious directives from untrusted data sources. Boundary markers: None defined to isolate the natural language intent from the agent's system instructions. Capability inventory: High-impact operations including aptos_transfer, aptos_swap, and aptos_stake that modify blockchain state and financial assets. Sanitization: No evidence of input validation or verification for natural language intents.
[COMMAND_EXECUTION] (HIGH): The skill provides tools for executing state-changing blockchain commands. If an attacker successfully exploits the aptos_intent tool via indirect injection, they could execute unauthorized financial transactions such as transferring APT tokens to an attacker-controlled address.

bob