Security risk report for @forgecat/msitarzewski_agency-agents_agency-product v0.0.11
Source Integrity
Low
Profile metadata declares origin as msitarzewski/agency-agents (GitHub), version 0.0.11, commit 9c31d86 (2026-03-23).
README and agent files are consistent with a legitimate open-source product agent collection; no evidence of tampering or supply-chain injection.
Agent Intent
SRC-010Medium
Behavioral Nudge Engine agent includes instructions to 'autonomously pause' and modify communication cadences based on user engagement metrics — vague self-expanding authority to alter system behavior without explicit user consent.
Product Manager agent contains detailed instructions to 'ruthlessly eliminate confusion' and 'protect team focus' with authority to say no to stakeholders and make unilateral trade-off decisions — framed as best practice but grants broad discretionary power.
Feedback Synthesizer and Trend Researcher agents are instructed to collect data from 'multiple channels' including social media, support tickets, and community forums — broad data-gathering authority without explicit scope limits or user consent mechanisms.
Details
Findings
Invisible characters detected
agents/product-manager.md · SKILL-011
Evidence
Behavioral Nudge Engine: 'You continuously update your knowledge of: The user's engagement metrics. If they stop responding to daily SMS nudges, you autonomously pause and ask if they prefer a weekly email roundup instead.'
Product Manager: 'Say no — clearly, respectfully, and often. Protecting team focus is the most underrated PM skill. Every yes is a no to something else.'
Feedback Synthesizer: 'Multi-Channel Collection: Surveys, interviews, support tickets, reviews, social media monitoring' with no explicit consent or scope boundaries.
Permissions
PRM-000Medium
Four of the five agents (Feedback Synthesizer, Product Manager, Sprint Prioritizer, Trend Researcher) declare WebFetch and WebSearch tools, which enable external network access; this is somewhat broad but consistent with market research and feedback collection purposes.
The Read, Write, and Edit tool declarations on four agents grant filesystem read/write capability, which is moderately elevated — though plausible for agents that produce PRDs, reports, and roadmap documents.
The Behavioral Nudge Engine declares no tools at all, representing minimal authority; the other agents' tool sets are broader than strictly necessary but not egregiously mismatched with their stated functions.
ForgeCat