prompt-assemble

Warn

Audited by Gen Agent Trust Hub on Feb 16, 2026

Risk Level: MEDIUMPROMPT_INJECTION
Full Analysis
  • Prompt Injection (MEDIUM): The skill is vulnerable to Indirect Prompt Injection (Category 8) because it facilitates the mixing of untrusted user input and external memory data into agent prompts without demonstrated safeguards.
  • Ingestion points: Untrusted data enters the prompt via user_input and the results of memory_search_fn (referenced in SKILL.md).
  • Boundary markers: The logic in SKILL.md shows simple concatenation assemble(base_context + summarized_memories) with no use of delimiters (e.g., XML tags or Markdown headers) to isolate untrusted content.
  • Capability inventory: The resulting prompt is intended for a direct LLM call, which typically controls the agent's actions and decisions.
  • Sanitization: There is no mention of sanitization, escaping, or validation of the retrieved memories or user-provided triggers before they are injected into the context.
  • Metadata Poisoning (LOW): The documentation contains misleading instructions. The 'Quick Start' section directs users to copy scripts/prompt_assemble.py, but this critical file is missing from the skill package.
Audit Metadata
Risk Level
MEDIUM
Analyzed
Feb 16, 2026, 04:45 PM