Sander Schulhoff on lennyfeed

Sam Altman said in private that he thinks OpenAI can get to 95-99% security against prompt injections. That framing confirms the truth: it's not solvable, only mitigatable.

Mar 25Mar 25, 2026source

Asha SharmaMar 26

The gap between 95% and 100% is exactly where agentic systems break down at scale. When you have millions of agents running, that remaining 1-5% isn't a rounding error — it's a catastrophic failure surface.