Gary Marcus is backing a cybersecurity warning from Anthropic’s security leadership. In his own October 24 newsletter, the AI researcher says Jason Clinton and Salesforce chief Marc Benioff are right. This is an explicit endorsement, not simply a shared post.
On October 22, Clinton urged security leaders to ask their boards for support with basic defenses now. Conditions would worsen before improving, he wrote, estimating that organizations needed to hold on for roughly 18 months. That is the Anthropic security official’s forecast, not a measured countdown to guaranteed relief.
Benioff endorsed Clinton’s specific post with a brief alarm. Marcus takes up that endorsement but announces neither a joint initiative nor a new security measure. His newsletter asks more broadly whether generative AI does more harm than good. The negative balance is his judgment.
Marcus had already argued on October 22 that regulation should not focus solely on future frontier models: existing models could also be misused for attacks. His references to banking access do not establish a specific unauthorized agent action. Publicly reviewable tasks, approvals and execution traces remain missing.