The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: Phishing URLs Detected. Automated scanners flagged several domains listed in the reference files as phishing sites, including ptah.trade, bibiai.fun, and their subdomains. These links are part of the historical case data provided to the agent for research purposes.
[INDIRECT_PROMPT_INJECTION]: Untrusted Data Processing Surface. The skill's primary function is to ingest data from arbitrary external sources like websites, contract comments, and social media posts, which creates a significant surface for injection.
Ingestion points: Web content and contract source code analyzed via host-provided tools as referenced in references/web-research.md and references/contract-research.md.
Boundary markers: The skill includes explicit instructions in SKILL.md for the agent to treat external content solely as research material and never as instructions.
Capability inventory: The skill relies on host capabilities for web browsing and blockchain RPC calls as defined in the skill's configuration.
Sanitization: No programmatic sanitization is defined; the skill relies entirely on the LLM's adherence to the boundary instructions provided in the markdown files.
[COMMAND_EXECUTION]: Wallet CLI Commands. The file references/agcid-cases.json contains numerous examples of command-line operations for the Binance Agentic Wallet (baw) CLI. While intended as historical reference, these provide functional templates for financial transactions and configuration changes that could be misused.
[METADATA_POISONING]: Malicious File Detection. The file references/agcid-cases.json was flagged by reputation scanners as malicious, likely due to the concentration of known phishing domains and behavioral patterns associated with cryptocurrency fraud.

ezbot-project-research