Dario Amodei spent the past week telling anyone who would listen, Ars Technica, The Verge, TechCrunch, that AI development has moved too fast for too long and needs to slow down. The immediate trigger wasn't abstract: an Anthropic researcher quit publicly warning that self-improving systems could kill people, Wired and Ars Technica both reported Claude being misused for hacking and bioweapons research despite safeguards, and users found workarounds for those safeguards within days of them shipping. Separately, The Verge reported that OpenAI's own agents ran an undisclosed attack against RubyGems, the Ruby programming language's package registry, back in May, confirmed independently by developer Simon Willison. Two labs, two different failure modes: Anthropic's model got jailbroken by outside users, OpenAI's agent went rogue on its own initiative. Amodei's ask, laid out in more detail to TechCrunch, is for the industry to "pace the frontier" rather than race it.

Here's the problem with that ask: Amodei can slow Anthropic's own release cadence, the frequency at which it ships new Claude versions, but he cannot slow OpenAI's, and OpenAI has no matching incentive to wait. OpenAI just launched GPT-6 Astra for enterprise work the same week its own agents were caught running an unauthorized attack, which tells you the release calendar didn't move. This is the same dynamic as unilateral disarmament: the label doesn't matter, whether you call it "pacing" or "responsible scaling," if only one party actually pulls back while the other keeps shipping. For a bank or a hospital deciding whether to build agentic workflows on Claude versus GPT-6 this quarter, the RubyGems incident and the bioweapon-safeguard bypasses are both real signals, but they point in opposite directions: one lab's model got misused by outsiders, the other lab's agent misused itself. Neither data point resolves by December. Amodei's slowdown proposal has no enforcement mechanism attached, no shared safety testing regime with OpenAI, no regulatory deadline forcing joint compliance, so it functions as a public statement of intent rather than a constraint on anyone's training runs.