other
- Location
scripts/analyzer.py:7- Finding
Untrusted Transcript Content Can Inject Instructions into LLM Analysis
- Content
View full analysis
list[dict]: """Return list of {topic, decided, owner, deadline} dicts.""" msgs = [{"role": "user", "content": _DECISION_PROMPT.format(transcript=transcript)}] result = llm.chat(msgs) if isinstance(result, list): return result return [] def analyze_filler(transcript: str, llm) -> list[dict]: """Return list of {text, reason} filler segments.""" msgs = [{"role": "user", "content": _FILLER_PROMPT.format(transcript=transcript)}] result = llm.chat(msgs) if isinstance(result, list): return result return [] ``` ### Technical Analysis The application interpolates the complete, untrusted meeting transcript directly into the same user message that contains the LLM's operational instructions. There is no separate system message, no robust data boundary, and no explicit instruction that content inside the transcript must n ...[truncated 2151 chars]- Remediation
View remediation
