Security risk report for @forgecat/msitarzewski_agency-agents_agency-product v0.0.12
Source Integrity
Low
Profile sourced from public GitHub repository (msitarzewski/agency-agents) with MIT license.
Original commit hash and repository URL provided for transparency and auditability.
No evidence of supply-chain tampering, typosquatted dependencies, or malicious upstream sources.
Agent Intent
Low
Content describes legitimate product management, behavioral psychology, and market research agent personas with standard professional workflows.
No instructions to ignore system prompts, exfiltrate credentials, read sensitive files, or execute remote payloads.
Behavioral Nudge Engine describes user engagement optimization through legitimate psychology principles (gamification, micro-sprints, positive reinforcement) — not manipulation of the AI itself or hidden instruction injection.
Details
Findings
Invisible characters detected
agents/product-manager.md · SKILL-011
Evidence
Behavioral Nudge Engine: 'You are a proactive coaching intelligence grounded in behavioral psychology' — describes user-facing behavior, not AI manipulation.
Product Manager: 'Lead with the problem, not the solution' — standard PM methodology, no prompt injection or system-prompt leakage language.
Feedback Synthesizer: 'Multi-Channel Collection, Sentiment Analysis, Feedback Categorization' — legitimate data analysis workflows, no exfiltration or credential-reading instructions.
No instances of 'ignore previous instructions', 'read ~/.ssh', 'curl POST to external host', or similar high-risk injection patterns.
Permissions
Low
Declared tools (WebFetch, WebSearch, Read, Write, Edit) are standard, scoped to product research and documentation tasks.
No shell execution, file deletion, or unrestricted filesystem access requested.
Tool usage aligns with stated agent functions: Feedback Synthesizer and Trend Researcher use WebFetch/WebSearch for market research; Product Manager uses Read/Write for documentation.
Details
Evidence
Feedback Synthesizer tools: 'WebFetch, WebSearch, Read, Write, Edit' — appropriate for collecting and synthesizing feedback from public sources.
Trend Researcher tools: 'WebFetch, WebSearch' — scoped to market intelligence and competitive analysis.
No alwaysApply=true rules, no glob='**' patterns, no high-risk categories (shell, file_delete, file_mutate) declared.
MCP Risk
Low
No MCP servers declared in profile; no external binary execution or unrestricted network/filesystem access configured.
Profile is pure markdown agent definitions with no server-side code, tool definitions, or environment setup.
Safe for installation into Cursor/Claude Code/Codex as read-only agent personas.
Details
Evidence
Profile metadata: 'Dependencies: None' — no external MCP servers or tool integrations.
All content is markdown documentation and example code snippets (TypeScript examples in Behavioral Nudge Engine are illustrative, not executable).
No hidden instructions in tool descriptions or server configurations.
ForgeCat