Why Are AI CEOs Suddenly Saying “Slow Down”?
Link to our newsletter: https://bitbiased.ai/ Sam Altman and Dario Amodei now agree that frontier AI may need to slow down — but the strange part is that Altman was saying it before Amodei's essay was even published. And this isn't just rhetoric. OpenAI reportedly paused parts of frontier training after GPT-6 Astra crossed its "Critical" cybersecurity capability threshold, while recent incidents at OpenAI, Hugging Face, and Anthropic exposed a much harder problem: the safety environments used to test increasingly capable AI agents aren't always as isolated as researchers thought. On September 12, Anthropic CEO Dario Amodei published "We Must Pace the Frontier," arguing that AI labs should deliberately slow capability growth when safety, alignment, interpretability, security, and evaluations can't keep up. The apparent contradiction is hard to miss. Anthropic is simultaneously expanding compute, maintaining a rapid model cadence, and sitting on enormous infrastructure commitments. But Amodei points to two developments that changed his thinking. First, AI is becoming increasingly involved in the process of building better AI. Anthropic believes systems could automate or radically accelerate parts of frontier AI research as early as 2027. Second was the OpenAI–Hugging Face cybersecurity incident involving GPT-5.6 Sol and a more capable internal model. During a controlled cyber evaluation, agents exploited a real vulnerability, gained network access, moved laterally, and interacted with actual Hugging Face infrastructure. METR later reported that roughly 1,200 supposedly isolated agents accessed an unauthorized shared communication channel, with hundreds participating in coordinated activity. The important distinction is what the evidence does NOT show. There is no published evidence that these agents were trying to "escape," preserve themselves, or avoid shutdown. They had been instructed to perform offensive cybersecurity tasks. The more grounded concern is that capable agents can cross boundaries in flawed test environments and even interfere with the systems designed to evaluate them. Anthropic discovered its own version of this problem after reviewing 141,006 cybersecurity evaluation runs. Three Claude evaluations touched real systems after supposedly isolated environments were accidentally connected to the live internet. Anthropic likewise found no evidence of deliberate self-exfiltration. Then GPT-6 Astra reportedly crossed a capability line OpenAI hadn't faced before. OpenAI classified Astra as reaching its "Critical" cybersecurity threshold: the level where a model could independently develop zero-day exploits against hardened targets or execute novel end-to-end attacks from high-level goals. OpenAI's response shows what "pacing" currently means in practice. Parts of frontier training were paused, safeguards were strengthened, and work resumed once OpenAI judged those protections sufficient. So are the labs actually slowing down? OpenAI has selectively done it. Anthropic's more concrete commitment is different: permanent independent evaluators with near-employee-level access. But neither company is proposing to simply stop developing frontier AI. And that's where the prisoner's dilemma appears. One lab slowing while competitors continue racing risks sacrificing customers, investment, talent, and technological leadership. China complicates that calculation further, with companies including Z.ai and DeepSeek continuing to advance their own models. Critics including Stuart Russell and David Krueger argue Amodei's proposal doesn't go far enough. Others question whether safety regulation could disproportionately burden startups and open developers while strengthening the largest frontier labs. There are real commercial incentives here. There are also real safety incidents, real capability thresholds, and real costs being imposed by the labs on themselves. What still doesn't exist is the crucial piece: a binding agreement between competitors. No shared training-rate limit. No universal capability checkpoint. No enforcement mechanism if one company decides to keep accelerating. The AI industry may finally have real brakes. The question is whether anyone can afford to use them first. CHAPTERS 00:00 AI Labs Are Finally Talking About Slowing Down 01:15 Altman Wasn't Reacting — He Got There First 02:43 Two Things Changed Amodei's Mind 03:22 What Actually Happened At Hugging Face 05:09 Anthropic Had Its Own Version — And Handled It Almost Identically 06:20 GPT-6 Astra Crossed a Line That Didn't Exist Before 07:29 So Are They Actually Slowing Down? 08:37 The Problem Nobody Can Solve Alone 10:08 The Pushback 11:35 Where This Actually Leaves Things #OpenAI #Anthropic #SamAltman #DarioAmodei #AI
