launch-monitor
Pass
Audited by Gen Agent Trust Hub on Aug 20, 2026
Risk Level: SAFE
Full Analysis
- [COMMAND_EXECUTION]: The skill uses standard CLI tools (
grep,date,cat,mkdir,python3,jq,ls) via the allowed-tools list to manage its internal reporting structure and state in~/.nimble/. These operations are well-scoped and align with typical agentic workflows. - [EXTERNAL_DOWNLOADS]: The skill fetches public marketing and news data from well-known press outlets (e.g., TechCrunch, Wired) and community platforms (e.g., Reddit, Hacker News) through the Nimble search infrastructure. These are documented as primary functional sources and do not include remote code execution.
- [DATA_EXPOSURE_&_EXFILTRATION]: The skill reads and writes to
~/.nimble/business-profile.jsonand~/.nimble/memory/. This data storage is local to the user's environment and used for legitimate persistence of preferences and reports. No hardcoded credentials or unauthorized exfiltration to third-party servers were found. - [PROMPT_INJECTION]: The instructions contain natural organizational markers like 'IMPORTANT' and 'REQUIRED' to guide the agent's logic, but these are for structural consistency (e.g., output templates) rather than attempting to bypass safety filters or override system prompts.
- [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted data from the web (press/social signals). It implements defensive boundary markers by instructing the agent to triage signals by urgency and materiality threshold, and it uses a fixed HTML template to safely interpolate data for user review, reducing the risk of data influencing the agent's internal state.
Audit Metadata