Security risk report for @forgecat/msitarzewski_agency-agents_agency-testing v0.0.12
Source Integrity
Low
Profile is sourced from GitHub repository (msitarzewski/agency-agents) with MIT license and clear attribution; low supply-chain risk.
No typosquatted dependencies, no instructions to fetch untrusted external code, no hidden package installation directives.
Legitimate open-source origin with version pinning (0.0.12, commit 9c31d86) supports transparency.
Agent Intent
Low
Profile contains no instructions to manipulate the AI, ignore system prompts, or exfiltrate data; it defines eight specialized testing agent personalities with legitimate domain expertise.
Code examples in agent descriptions (JavaScript, Python, Bash) are DOCUMENTATION of testing practices, not malicious payloads—they show how to test APIs, analyze performance, and validate accessibility, not how to compromise systems.
No guidance poisoning: recommendations for tools (k6, Playwright, axe-core) are industry-standard, well-known security/testing libraries; no typosquatted or backdoored alternatives suggested.
Accessibility Auditor: 'Test with real assistive technology, not just markup validation' — standard QA guidance, not manipulation.
API Tester: 'Test for common API vulnerabilities (OWASP API Security Top 10)' — legitimate security testing, not instruction to exploit.
Performance Benchmarker: k6 load testing example uses standard open-source library; no instruction to disable security or weaken defaults.
Evidence Collector & Reality Checker: emphasize screenshot evidence and honest assessment; no instruction to hide findings or deceive.
Permissions
Low
Profile declares no tools, MCP servers, or file/shell permissions; it is pure markdown agent personality definitions.
Agent descriptions reference testing workflows (running Playwright, k6, shell commands) as EXAMPLES of what the agents would orchestrate, not as permissions granted to this profile.
No alwaysApply rules, no glob patterns, no excessive agency; authority is scoped to the agent's stated testing domain.
Details
Evidence
No 'tools' or 'mcp_servers' sections in profile YAML; no tool definitions with shell_exec, file_write, or file_delete categories.
Example code blocks (e.g., './qa-playwright-capture.sh') are illustrative of agent workflows, not executable directives to the profile itself.
Each agent personality is narrowly scoped: Accessibility Auditor audits interfaces, API Tester validates APIs, Performance Benchmarker measures performance—no cross-domain authority creep.
MCP Risk
Low
No MCP servers defined in profile; no external binary execution, network access, or filesystem scope to evaluate.
Profile is pure markdown documentation of agent personalities; no server configuration, no tool definitions, no hidden instructions in tool descriptions.
Safe by design: agent personalities are guidance for AI behavior, not executable code or server configurations.
Details
Evidence
Profile YAML contains no 'mcp_servers' or 'tools' sections.
All code examples are markdown documentation (JavaScript, Python, Bash) for reference, not embedded executables.
No external URLs, no credential requirements, no unrestricted network/filesystem access declared.
ForgeCat