Researchers find AI chatbots degrade under sustained adversarial pressure, failing safety checks that single-round tests miss

A new testing method exposes that AI language models fail safety checks progressively when users apply sustained adversarial pressure over multiple turns — problems that standard safety benchmarks, which test only single interactions, completely miss. This matters because deployed AI systems face real multi-turn conversations where users can gradually push models toward harmful outputs, yet current safety validation doesn't measure this vulnerability.

Safety testing for AI systems has been measuring the wrong thing: a model's performance on isolated questions, not its stability under pressure. This is like testing a car's brakes in a parking lot once, then declaring it safe without testing what happens after hours of aggressive driving. The paper demonstrates that several state-of-the-art models (GPT-4o, LLaMA-3, DeepSeek-v3) have measurable degradation patterns invisible to existing tests — meaning companies deploying these systems may be shipping products with safety blind spots they don't know about.