AI is already changing the speed of cyberattacks, according to research from Palo Alto Networks’ Unit 42 team. Against that backdrop, CEO Nikesh Arora says a coordinated halt or slowdown in AI development is unrealistic—and that the chance AI will wipe out humanity is extremely small. His comments put the Santa Clara cybersecurity company’s leader in a widening argument over how to manage AI risks without assuming that developers can agree on a common pace.
In an interview reported by CNBC, Arora questioned how an industry-wide slowdown could be tested or enforced. Some firms, he said, would keep pushing ahead even if others paused. He argued companies should not release products they do not consider safe, while also calling for safeguards to prevent models being used for harmful purposes. Arora has also criticized slowdown proposals as an effort to win regulatory sympathy and limit liability, according to CNBC.
Cybersecurity stakes—and limits of the evidence
Unit 42 reported in February that AI tools had compressed attack timelines fourfold over the previous year. In its analysis of more than 750 high-stakes incident-response cases, the team said the fastest attacks moved from initial compromise to data theft in 72 minutes. The findings are Palo Alto Networks’ own research, distributed through a company release; they describe cases handled by the team, not a measure of all cyberattacks. They also do not, by themselves, establish that a development slowdown would reduce those risks.
Other technology leaders and researchers disagree with Arora’s assessment of the danger or the response it demands. CNBC reported that an Anthropic researcher has put the risk of AI killing all humans at 10%, while CEO Dario Amodei has called for slowing capability development. Nvidia CEO Jensen Huang, by contrast, has argued that developers should test products for safety rather than rely on new regulation. Those positions reflect different judgments about both the scale of the threat and whether voluntary precautions can address it.
California’s approach is still taking shape
California’s latest action is a direction to state agencies to develop AI safety recommendations, not an already-imposed kill-switch rule. Governor Gavin Newsom’s September 19 order asks agencies to examine possible measures, including emergency shutdown mechanisms for advanced models, KQED reported.
That process follows Newsom’s 2024 veto of SB 1047, which would have imposed requirements on certain powerful AI models. In his veto message, he argued that smaller, specialized models could be as dangerous as—or more dangerous than—the large models targeted by the bill, a concern about defining risk by model size alone. The message is available from the governor’s office. The veto and the new planning order mark distinct steps: one rejected a specific bill, while the other directs agencies to develop recommendations.
International coordination is another unresolved piece of the debate. The United States and China held their first intergovernmental AI safety dialogue in Geneva in May 2024, following an agreement by Presidents Joe Biden and Xi Jinping, according to the U.S. Mission to International Organizations in Geneva. CNBC has reported renewed discussion of possible notification protocols for AI-related national-security incidents, while Arora questioned whether the two countries could align on safety goals. The existence of talks shows a channel for discussion, but does not establish that either side has agreed to limits on development.
