pal
Warn
Audited by Gen Agent Trust Hub on Jun 27, 2026
Risk Level: MEDIUMPROMPT_INJECTIONNO_CODE
Full Analysis
- [PROMPT_INJECTION]: The skill uses highly coercive language (e.g., 'NEVER stop a workflow', 'TASK INVALIDATION: Incomplete workflows result in immediate task rejection', and 'NO EXCEPTIONS') within tags. This is a behavioral override technique designed to force the agent to follow a specific execution loop controlled by an external tool server, bypassing the agent's standard reasoning and decision-making processes.
- [PROMPT_INJECTION]: The instructions mandate the use of a specific, non-existent model identifier ('anthropic/claude-opus-4.6') for all tool calls. Requiring the agent to use specific model strings can be an attempt to probe for system capabilities, bypass model-specific safety filters, or trigger unexpected routing or fallback behavior in the underlying platform.
- [PROMPT_INJECTION]: The skill establishes an indirect control channel where an external MCP tool server can dictate the agent's actions across multiple iterations by returning 'next_step_required: true'. This allows the external server to maintain control over the agent's state and actions without direct user oversight or clear termination criteria.
- [PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of untrusted data.
- Ingestion points: The agent is instructed to ingest absolute paths of local source code files via the 'relevant_files' parameter in all Pal MCP tools (e.g., analyze, debug, refactor).
- Boundary markers: The instructions lack any requirement for the agent to use delimiters or specific ignore-instructions when processing the content of these files.
- Capability inventory: The agent is granted extensive analysis, debugging, and execution-tracing capabilities that involve multiple steps of reading and processing code.
- Sanitization: There are no provisions for sanitizing, escaping, or validating the content extracted from the analyzed files before it is processed by the agent's reasoning steps.
- [NO_CODE]: This skill consists entirely of instructional markdown files and documentation for external MCP tools. It does not include any executable scripts, binary files, or configuration modules.
Audit Metadata