atlas-weave

Pass

Audited by Gen Agent Trust Hub on Nov 22, 2026

Risk Level: SAFEINDIRECT_PROMPT_INJECTIONCOMMAND_EXECUTION
Full Analysis
  • [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted user input via the primary argument-hint to define coding tasks. This input is then propagated to multiple sub-agents (such as the planner and coder) which have the capability to modify files and execute shell commands.
  • Ingestion points: The user's "rough multi-file coding request" enters the agent context at the DECOMPOSED stage and is stored in a task packet.
  • Boundary markers: The instructions specify that the intent is frozen into an immutable task packet to prevent runtime manipulation of the goal, though the content itself remains untrusted input.
  • Capability inventory: The skill and its sub-runs utilize high-privilege tools including Write, Edit, and Bash.
  • Sanitization: No explicit sanitization, escaping, or validation of the user-provided request is described before it is interpolated into prompts for the planning and coding agents.
  • [COMMAND_EXECUTION]: The orchestrator performs system command execution using the Bash tool for several management and verification tasks.
  • It executes free -m to monitor system memory availability, which is used as a live admission input for wave scheduling.
  • It executes a user-provided verify_cmd to validate changes on the merged worktree during the INTEGRATE stage.
  • It executes local Python modules via python3 -c using a configured PYTHONPATH that points to directories relative to the skill directory.
Audit Metadata
Risk Level
SAFE
Analyzed
Nov 22, 2026, 01:11 AM