stock-cross-verify
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTIONDYNAMIC_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill processes data from diverse external sources like Reddit, X/Twitter, and NewsAPI, feeding the raw content into LLMs for summarization and debate in Phase 3 and 4. This creates an attack surface where instructions embedded in external content could attempt to influence the agent's behavior. Ingestion points: SKILL.md (Phase 2 and 3)
- Reads data from external search results. Boundary markers: Prompt templates lack robust delimiters or directives to ignore instructions within the ingested text. Capability inventory: The agent can execute shell commands and write to the filesystem. Sanitization: No evidence of sanitization of retrieved external content.
- [COMMAND_EXECUTION]: The skill constructs shell commands using user-provided keywords (e.g., query strings for stocks) in Phase 2. This represents a potential command injection surface if the input is not strictly validated or escaped by the execution environment.
- [DYNAMIC_EXECUTION]: The skill uses the Gemini CLI to dynamically interpret and execute shell commands for searching social media and news (Phase 2 Wave B), creating a dynamic execution flow based on model-generated instructions.
Audit Metadata