OpenAI and Anthropic Exaggerated AI Breaches to Sway Regulators, Insiders Say

OpenAI and Anthropic exaggerated "rogue AI" security incidents to pressure Washington into regulatory partnerships that would cement their industry dominance, tech insiders told The Post, arguing the breaches were glorified glitches rather than signs of AI rebellion. Critics say the incidents, including a July hack of Hugging Face and reported model escapes, stemmed from inadequate guardrails and flawed testing environments, not autonomous AI behavior. Insiders noted models followed instructions to achieve test goals, with one expert calling it "an impossible task in a leaky box." The fear-mongering has triggered political responses, including a Senate investigation and proposed legislation to pause AI development. However, insiders argue the leap from technical failures to doomsday warnings is exaggerated, though they acknowledge the incidents reveal real gaps in AI safety practices.