The Agent Skills Directory

[CREDENTIALS_UNSAFE]: A hardcoded Uniswap API key (NeoYO3V50_koJAipDEalYWbMO1XMaFPAQmpOm6_Npo0) is present in plugins/uniswap.md. Hardcoding credentials is a dangerous practice that can lead to unauthorized access and credit exhaustion.
[REMOTE_CODE_EXECUTION]: The skill facilitates the execution of remote code from unverified sources. plugins/aerodrome.md uses uvx to download and run the Sugar SDK directly from a non-standard GitHub repository (velodrome-finance/sugar-sdk.git). Similarly, plugins/morpho.md uses npx to execute the Morpho CLI from the @morpho-org/cli package. These sources are not among the verified providers for this platform.
[COMMAND_EXECUTION]: The instructions in references/approval-mode.md direct the agent to automatically execute shell commands (open, xdg-open, start, Start-Process) to open URLs in the user's default browser. If the agent is influenced by malicious input, this capability could be used to force the user to visit dangerous websites or potentially execute local commands if URL schemes are exploited.
[INDIRECT_PROMPT_INJECTION]: The skill interacts with numerous third-party DeFi and discovery APIs (Uniswap, Morpho, Moonwell, Avantis, Bankr). The agent ingests data from these external endpoints (such as token names, metadata, and position details) and uses it to construct and execute onchain transactions via the send_calls tool. This creates a surface for indirect prompt injection where a malicious API response could manipulate transaction parameters.
[EXTERNAL_DOWNLOADS]: The skill is configured to fetch fallback documentation and configuration files from docs.base.org using the web_request tool whenever local files are not accessible. While targeting a well-known service, this involves dynamic fetching of instructional content that modifies agent behavior at runtime.

base-mcp