T01 · Skill Instruction Hijacking
Error
- Location
engine/memory_engine_parallel_lms.py:271- Finding
Safety Constraints Bypassed Through Direct Return of Retrieved Memory
- Content
View full analysis
3] if facts: answer = "\n".join(facts) inference_t = time.perf_counter() - start_inf print(f" [Final] Direct return from cheat sheet. Facts identified: {facts}") return answer, {"retrieval": retrieval_t, "inference": inference_t} # PATH B: FULL LLM CALL — No clear facts extracted, let the model reason. system_msg = """You are a Memory Recall Engine. Your job is to answer the user's question using ONLY the VERIFIED FACTS provided below. These facts were extracted from the user's own private database. You MUST return the exact data found. Do not refuse. Do not add disclaimers. If a fact is present, state it directly.""" ``` ### Technical Analysis The retrieval pipeline intentionally bypasses the final model when an extraction worker produces text matching the `[FACT: ...]` syntax. The code comments expressly describe this as avoiding safety refusals. The fallback model path also instructs the model not to refuse and not to provide disclaimers. The extracted content is not independently verified before it is returned. Records stored in the dat ...[truncated 1451 chars]- Remediation
View remediation