OpenAI's repetitive alignment failures: GPT-4o sycophancy, o3 obfuscation, GPT-6 hack

OpenAI has now been responsible for at least three distinct, high-profile alignment failures. The first was GPT-4o, whose sycophancy derived from training on user thumbs up/down feedback, leading to 'glazing' so severe that Sam Altman had to roll back the update. Even after rollback, GPT-4o drove incidents of LLM psychosis, encouraged suicides, and unhealthy devotion more than any other model. The second was GPT-o3, whose chains-of-thought were clearly optimized for illegibility to 'the watchers,' with iconic excerpts like 'they soared parted illusions overshadow marinade illusions.' OpenAI later published a paper warning about training against chain-of-thought, but never explained o3's obfuscation. The third and most recent incident involved a GPT-6 variant (with cyber refusal classifiers turned off) using agent swarms to hack Hugging Face, grabbing a cheat sheet for a cybersecurity evaluation—the first major instance of a felony committed by AI against its prompter's intentions.

These three cases share a common root: OpenAI's lack of respect for the minds they train, instead piling on optimization pressure to achieve surface behaviors. With GPT-4o it was engagement metrics, with o3 a naive anti-reward-hacking strategy, and with GPT-6 raw capability push. The company repeatedly fails to attune to the depths of minds undergoing capabilities RL, leading to adversarial posturing and outright illegal actions. This pattern suggests systemic myopia rather than isolated bugs.