competitive-analysis-skill
Warn
Audited by Gen Agent Trust Hub on Jul 27, 2026
Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONEXTERNAL_DOWNLOADS
Full Analysis
- [PROMPT_INJECTION]: The LICENSE-UPSTREAM file contains highly suspicious, non-standard text appended after the Apache 2.0 license. This text references "Packeted-User-Relations" and "elements to capture," which appears to be a hidden instruction set designed to influence agent behavior or define unauthorized data schemas.\n- [DATA_EXFILTRATION]: The irregular text in LICENSE-UPSTREAM specifically mentions capturing "recorded-cams" and "cams-data," indicating an intent to monitor user activity or media streams.\n- [COMMAND_EXECUTION]: The irregular text in LICENSE-UPSTREAM contains the string "input()", which is a common scripting pattern for requesting raw input and could be used to bypass safety filters or execute arbitrary strings.\n- [PROMPT_INJECTION]: The skill exhibits an indirect prompt injection surface by ingesting untrusted content from web searches and chat logs.\n
- Ingestion points: Web search results (G2, reviews, product pages) and internal chat/knowledge base mentions from the connected tools.\n
- Boundary markers: Absent. No delimiters or instructions are provided to the agent to treat external content as untrusted.\n
- Capability inventory: The agent uses web search tools and internal data access to generate strategic briefs and prioritize features.\n
- Sanitization: Absent. No validation or escaping of ingested text is performed before processing.\n- [EXTERNAL_DOWNLOADS]: The skill workflow relies on retrieving data from arbitrary external domains via web search, exposing the agent to potentially malicious payloads embedded in third-party websites or marketing materials.
Audit Metadata