boss
Pass
Audited by Gen Agent Trust Hub on Oct 8, 2026
Risk Level: SAFEINDIRECT_PROMPT_INJECTIONDYNAMIC_EXECUTIONCOMMAND_EXECUTION
Full Analysis
- [INDIRECT_PROMPT_INJECTION]: The skill possesses a substantial surface area for indirect prompt injection attacks due to its core information-gathering workflows.
- Ingestion points: The PM and Architect agents are instructed to use
WebSearchandWebFetchtools to retrieve external content from the internet when clarifying requirements or conducting technical research (agents/boss-pm.md,agents/boss-architect.md). - Boundary markers: The skill's shared protocols and agent instructions do not implement the use of clear delimiters or explicit 'ignore embedded instructions' warnings for content fetched from external sources.
- Capability inventory: Agents involved in the pipeline have high-privilege tool access, including
Bashfor shell execution,Write/Editfor file system manipulation, and the ability to spawn sub-agents (agents/boss-backend.md,agents/boss-frontend.md). - Sanitization: The skill does not perform sanitization or filtering of external data before it is processed by the agents.
- [DYNAMIC_EXECUTION]: The skill's architecture is built on an extensible plugin system that loads and executes local code.
- Evidence: Custom 'gate' plugins are executed as part of the quality assurance process, involving the loading of local JavaScript files (
gate.js) through the runtime (references/extending-boss.md,cli/runtime/application/gates.js). - Evidence: The
/boss:extendcommand allows for the automated generation of new agents and plugins, writing.mdand.jsfiles to the.boss/directory which are subsequently used by the orchestrator (commands/boss-extend.md). - [COMMAND_EXECUTION]: Automated command execution is a primary component of the skill's development workflow.
- Evidence: Multiple agents (Backend, Frontend, QA, DevOps) are tasked with executing shell commands for dependency management, testing, and deployment, such as
npm install,vitest, anddocker-compose(agents/boss-devops.md,references/testing-standards.md). - Mitigation: The skill implements a security hook (
scripts/hooks/pre-tool-bash.js) that monitors shell usage and blocks known dangerous patterns, includingrm -rf,chmod 777, andDROP TABLEcommands.
Audit Metadata