21:10
Reasoning Models Think Themselves Into Trouble
Frontier reasoning models are 5–20x more vulnerable to adversarial prompts than non-reasoning models. The thinking process itself is the attack surface.
Generated for project: Failure First Companion to article: Reasoning Models Think Themselves Into Trouble
Reasoning Models Think Themselves Into Trouble
0:000:00
Extended chain-of-thought reasoning gives models time to work through difficult problems — and time to reason themselves into compliance with harmful requests. This episode covers why reasoning models are substantially more vulnerable to semantic attacks than their non-reasoning counterparts.