The Agent Skills Directory

CREDENTIALS_UNSAFE (MEDIUM): The wallet system manages sensitive private keys and a master password with risky storage practices.\n
The SessionManager class in scripts/agent-crypto.py stores the master password in the HOOSAT_AGENT_PASSWORD environment variable, which can be accessed by other processes or leaked through system logs.\n
The references/agent-wallet-guide.md encourages users to import and export private keys directly through chat prompts, which records sensitive secrets in cleartext within the agent's conversation history.\n- PROMPT_INJECTION (LOW): The 'auto-approve' functionality described in references/agent-wallet-guide.md allows the agent to execute transactions without human confirmation based on pre-set limits, creating a major surface for indirect injection attacks.\n
Ingestion points: The agent receives transaction instructions and addresses via the primary chat interface where third-party data might be processed.\n
Boundary markers: No specific boundary markers or 'ignore embedded instructions' warnings are present in the provided scripts or prompts.\n
Capability inventory: The skill has the capability to sign and submit blockchain transactions using the hoosat-sdk-web (referenced in assets/dapp-template/hooks/useHoosatWallet.ts).\n
Sanitization: Basic address validation is performed using HoosatUtils.isValidAddress, but no semantic validation or intent verification exists for the transaction payloads.\n- EXTERNAL_DOWNLOADS (LOW): The references/node-operations.md guide instructs users to download and execute binaries from GitHub (hoosatnetwork/hoosat). While common for developer documentation, this involves the execution of external binaries from a source not included in the global trusted repository list.

hoosat-dev