start-here
Pass
Audited by Gen Agent Trust Hub on Nov 17, 2026
Risk Level: SAFECOMMAND_EXECUTIONDATA_EXFILTRATIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [COMMAND_EXECUTION]: The skill instructions and internal logic rely on the execution of local Python scripts to manage brand context, detect tool capabilities, and determine workflow routing.
- Evidence:
SKILL.mdandreferences/brand-memory.mdcontain instructions for the agent to run commands such aspython3 ${CLAUDE_SKILL_DIR}/scripts/brand_context.pyandpython3 ${CLAUDE_SKILL_DIR}/scripts/capability.pyto initialize brand context and check for available tools. - [DATA_EXFILTRATION]: The skill includes a
capability.pyscript designed to scan environment files (.env) for the presence of API keys (e.g.,FIRECRAWL_API_KEY,REPLICATE_API_TOKEN) to identify available integrations. Although the script is programmed not to print the values in its output, the retrieval and parsing of these keys from sensitive files into the script's memory constitute a data exposure surface. - Evidence:
scripts/capability.pyparses.envfiles to check for numerous sensitive API keys used by marketing toolsets and provides a manifest of which providers are available. - [INDIRECT_PROMPT_INJECTION]: The skill has a significant ingestion surface, as it fetches and processes content from the public web (websites, reviews, and search results) and reads user-authored files like
MARKETING.mdto build brand identities. This content could potentially contain malicious instructions intended to manipulate the agent. - Ingestion points: Web scraping of business homepages, services pages, and third-party review platforms (G2, Capterra, Reddit); reading user-authored
MARKETING.mdandpreferences.mdfiles. - Boundary markers: The skill employs data labels such as
Data: LIVE,Data: LOCAL DOCS, andData: MODEL KNOWLEDGEto categorize information sources and improve context awareness. - Capability inventory: The agent can execute local Python scripts via the shell, write files to the workspace (e.g., brand profiles, campaign assets), and perform network fetches.
- Sanitization: There is no explicit evidence of filtering or sanitizing fetched content to prevent instructions within the data from being executed by the LLM.
Audit Metadata