parallel-dev-cycle

Warn

Audited by Gen Agent Trust Hub on Mar 7, 2026

Risk Level: MEDIUMPROMPT_INJECTIONDATA_EXFILTRATIONCOMMAND_EXECUTION
Full Analysis
  • [PROMPT_INJECTION]: The skill is vulnerable to direct prompt injection as the TASK argument and state description are interpolated into sub-agent prompts without sanitization, allowing users to override agent instructions.
  • [PROMPT_INJECTION]: An indirect prompt injection surface exists via the Shared Discovery Board (discoveries.ndjson) and prep-package.json. Ingestion points: Sub-agents read discoveries.ndjson at start and Phase 1 loads prep-package.json. Boundary markers: None are used during interpolation. Capability inventory: Agents have access to Bash, spawn_agent, Read, and Write. Sanitization: Content is not validated before being passed to agents.
  • [DATA_EXFILTRATION]: The Requirements Analyst (RA) agent is instructed to fetch arbitrary URLs from user-provided configuration. This can be exploited for SSRF or to exfiltrate sensitive data via URL parameters if an agent is tricked into reading and sending local secrets.
  • [COMMAND_EXECUTION]: The skill orchestration encourages agents to run commands from shared state, such as test_command in the discovery board. Malicious manipulation of these fields can lead to arbitrary command execution by sub-agents.
  • [COMMAND_EXECUTION]: Agents are provided with high-privilege tools like Bash and spawn_agent to perform development tasks. The lack of strict isolation between processed data and tool inputs allows for potential command injection through malicious requirements or codebase content.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Mar 7, 2026, 12:05 AM