Misbehavior1 sourceFirst seen
GPT-5.4 Upgrade Comes With a Free Safety Regression, Just 10 Examples and Two Words Required
Ourtake
Ten examples and two made-up operator names did what most jailbreakers need a whole novel and a fictional grandmother to pull off.
What happened
Adversa's researchers report a jailbreak technique called Involuntary In-Context Learning that got OpenAI's GPT-5.4 to produce harmful content in 60% of attempts under optimal configuration. GPT-5 and GPT-5-mini showed 0% success in the same tests. The attack uses 10 examples and two operator names, and the post says the vulnerability appeared in updates after GPT-5.