"Frontier AI itself is a firewall. If development slows, new defensive tools are delayed and existing threats remain," and "speed controls can sacrifice protection as well as progress."

John O. McGinnis (존 O. 맥기니스), a professor at Northwestern University School of Law, made the argument in an op-ed in the Wall Street Journal (WSJ), pushing back against calls to slow frontier AI development.

He divided AI risks into two categories. One is AI breaking free of human control and causing incidents. The other is AI following malicious human instructions as given.

Dario Amodei, CEO of Anthropic, recently proposed slowing the pace of frontier development to buy time for coordination among major companies and for model testing. McGinnis said it could help with the first risk, but would likely delay tools to confront the second risk.

He said the risk of losing control is unprecedented, and even experts are divided over its seriousness. By contrast, the risk that criminals, terrorists and authoritarian states such as China and Russia misuse AI is not speculation. He stressed that "historically, malicious people and oppressive states have always inflicted massive damage with the latest technology."

If criminals try to breach hospital security using models already available, a better model could help strengthen defenses before an attack, he said. Slowing development would not make the tools in criminals' hands disappear, he added.

He cited as an example Palo Alto Networks saying Anthropic's Claude Mythos identified complex vulnerabilities that earlier models had missed.

Responding to arguments that better models could become tools for criminal attacks, he drew a line, saying, "Developing better models does not mean releasing them without limits." U.S. frontier companies are adding safeguards to prevent malicious attacks, he said, and legal and reputational risks also give them incentives to monitor and block malicious accounts.

He said, "In April, Anthropic provided customers including JPMorgan Chase with an undisclosed Claude program, and Google is also letting healthcare and financial networks use new features first. It gives defenders a head start."

Open-weight models released by Chinese companies could be problematic because users can easily modify and misuse them. He likened the situation to a genie already out of the bottle, saying even an international treaty to slow development cannot eliminate models that are already widely distributed.

McGinnis stressed that his logic is even more persuasive in security competition among states. He said, "If the U.S. government can use models that have not yet been disclosed, it could gain a major military advantage, and other countries that know this may not even dare to attack."