defi-portfolio-manager

Pass

Audited by Gen Agent Trust Hub on Aug 20, 2026

Risk Level: SAFEPROMPT_INJECTIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill processes untrusted content from several external sources including DeFi protocol web pages (via browser automation), Google Sheets, and crypto journalism platforms.
  • Ingestion points: Data enters the agent context through chrome-use (reading DeBank profiles), gws (reading Google Sheets), and the read-news pipeline.
  • Boundary markers: The instructions do not define explicit delimiters or 'ignore' instructions for the agent when interpolating content from these external sources into its decision-making prompt.
  • Capability inventory: The agent can execute local TypeScript scripts via bun, perform web searches, and utilize browser automation tools.
  • Sanitization: While the provided scripts include logic to validate blockchain-specific data (such as address format and canonical stablecoin addresses), there is no evidence of sanitization for natural language instructions that could be embedded in news articles or web page metadata.
  • [DYNAMIC_EXECUTION]: The skill relies on the execution of multiple localized helper scripts using the bun runtime to perform data validation and maintain a position cache.
  • Evidence: SKILL.md directs the agent to run portfolio_cache.ts, morpho_vault_status.ts, hyperliquid_status.ts, and merkl_rewards.ts. These scripts are included in the skill package and manage data logic, including reading and writing to a local CSV cache at .cache/defi-portfolio-manager/crypto-portfolio.csv.
  • [DATA_EXPOSURE_AND_EXFILTRATION]: Public wallet addresses provided by the user are transmitted to external DeFi services to retrieve balance and reward information.
  • Evidence: The skill's scripts perform HTTP requests to official API endpoints for Hyperliquid (api.hyperliquid.xyz), Merkl (api.merkl.xyz), and Morpho (api.morpho.org). This transmission is necessary for the skill's primary function and targets well-known, established services. No sensitive credentials (such as private keys or API tokens) are transmitted or stored by the skill.
  • [REMOTE_CODE_EXECUTION]: The skill makes network requests to fetch live market data from well-known DeFi aggregators.
  • Evidence: The skill uses curl and the fetch API to retrieve pool information from DefiLlama (yields.llama.fi) and other protocol-specific APIs. These requests are for data retrieval only and do not involve piping remote content to a shell or interpreter.
Audit Metadata
Risk Level
SAFE
Analyzed
Aug 20, 2026, 02:21 PM
Security Audit — agent-trust-hub — defi-portfolio-manager