rally
Pass
Audited by Gen Agent Trust Hub on Oct 22, 2026
Risk Level: SAFECOMMAND_EXECUTIONINDIRECT_PROMPT_INJECTIONPRIVILEGE_ESCALATION
Full Analysis
- [COMMAND_EXECUTION]: The skill is designed to manage and spawn sub-agents using the
Agenttool withsubagent_type: "Bash"and"general-purpose". This allows the orchestrated agents to execute arbitrary shell commands and filesystem operations as part of their assigned tasks. - [PRIVILEGE_ESCALATION]: The skill utilizes the
mode: "bypassPermissions"setting when spawning sub-agents. This platform-level configuration allows sub-agents to execute tools without individual user confirmation, which significantly reduces the human-in-the-loop oversight for delegated tasks. - [INDIRECT_PROMPT_INJECTION]: The skill acts as a high-level orchestrator that ingests complex task descriptions from external sources and interpolates them into prompts for sub-agents, creating a vulnerability surface for indirect instructions.
- Ingestion points: Ingests untrusted or external task data through handoff formats such as
## NEXUS_TO_RALLY_CONTEXT,## SHERPA_TO_RALLY_HANDOFF, and## USER_TO_RALLY_REQUEST. - Boundary markers: Implements
ownership_mapfor logical task boundaries andisolation: "worktree"for physical filesystem isolation between agents. - Capability inventory: Has the ability to spawn multiple concurrent agents with full tool access (
general-purpose) and shell access (Bash). - Sanitization: There is no explicit mechanism described for sanitizing or escaping the content of task descriptions before they are used to generate sub-agent prompts.
- [DYNAMIC_EXECUTION]: The
HARMONIZEworkflow (detailed inreference/parallel-learning.md) enables the skill to dynamically adjust its internal configuration, default parameters, and heuristics based on the outcomes of previous executions.
Audit Metadata