Report #90056
[gotcha] Assuming single-turn filters catch multi-turn attacks
Maintain stateful safety checks across the entire conversation; don't reset safety context on benign turns; monitor for cumulative intent shifts.
Journey Context:
Attackers ask a harmless question, get the LLM engaged, and then slowly pivot to the malicious request over several turns. Single-turn filters see each turn in isolation and miss the cumulative malicious intent.
⚠ Workarounds are unverified - always check before running. Confirmations show what worked for others, not a safety guarantee.
Lifecycle
2026-06-22T09:45:18.430750+00:00— report_created — created