AI Safety Systems Lag as OpenAI, Anthropic Clash

A Widening Gap Between Detection and Prevention

A new assessment of leading artificial intelligence labs finds that the industry has grown considerably more skilled at identifying risky model behavior than at actually stopping it, a distinction that safety researchers say matters more as models grow more capable and are deployed more widely 1. The finding lands at a moment when the two most prominent U.S. AI developers, OpenAI and Anthropic, are publicly diverging on how seriously to treat those risks — and on whether to slow down because of them.

OpenAI Pauses, Anthropic Holds Firm

The clearest sign of that divergence came this week when OpenAI said it was pausing certain model work over safety concerns, just days after Anthropic insisted its own safeguards were robust enough that it saw no need to change course 45. Coverage of the standoffframed OpenAI's move as effectively conceding ground in an argument the two companies have been having, at least implicitly, over how much caution the current moment demands 4. The disagreement is notable partly because it is playing out so publicly between two labs that are simultaneously racing each other commercially and, on other issues, presenting a united front on safety norms. It also raises the possibility that the companies could end up on different release schedules for future models, with one moving faster than the other depending on how each resolves its internal safety reviews 4.

A Regulatory Fight in Massachusetts

That commercial and philosophical rivalry has a policy dimension as well. OpenAI and Anthropic are also clashing over proposed AI safety legislation in Massachusetts, where the state Senate has passed an economic development bill containing some of the toughest AI oversight rules attempted at the state level 2. The measure would require AI labs to meet new safety and disclosure obligations, and the two companies appear to have staked out different positions on how burdensome those rules should be — a split that mirrors, in the regulatory arena, the same tension over speed versus caution playing out in their product decisions 2. With federal AI legislation largely stalled, state-level fights like Massachusetts' are increasingly where the practical rules of the road for AI companies are being written.

Risk and Reward in the Same Breakthrough

Underscoring how difficult it is to draw clean lines around AI risk, researchers recently used AI tools to help design a brand-new virus, a milestone that has alarmed some observers even as the scientists behind it argue the work could aid the fight against drug-resistant bacteria and ultimately save lives 3. The episode illustrates the broader challenge facing labs, regulators, and lawmakers alike: the same underlying capabilities that raise safety alarms often carry genuine scientific promise, making blanket caution or blanket permissiveness equally unsatisfying responses. Taken together, the reporting suggests an industry where technical safeguards, corporate strategy, and state-level regulation are all scrambling to keep pace with capabilities that are advancing faster than the systems meant to govern them.