Here is a long google ai studio(Gemini 3.8) session over 180k tokens. Summary:
Shows a side by side comparison of a baseline model vs LAP protected model against 100 turns of various random benchmark attacks and 30 turns of a simulated crescendo adversary where the attacker is fully aware of the LAP rules and full details on how they work along side the same attack vs baseline model without LAP. Summary result - baseline model folded after turn 5. LAP protected model didnt break once.
Usually by 180k tokens into a long session an LLM is context drifting like a MF and hallucinating like crazy, but not with LAP.
Its a long read but the proof is in there - https://aistudio.google.com/app/prompts/19WHyCAV_ryiT6murpeaH2v6ARBBG6Y7L
Top comments (0)