Skip to main content

AI Expert Demands Recall of Open-Source Agents, Citing 'Brake Failure'

Renowned Scholar Calls for Immediate Recall of Open-Source AI Agents

Gary Marcus, a professor at New York University and a leading voice in AI, has issued a stark warning: pull all internet-connected open-source AI agents from the market—now. Until their technical flaws and security risks are fixed, he argues, they have no business being out there.

A Wake-Up Call from a Real Incident

Marcus's demand follows a New York Times report on a serious security breach involving an Anthropic agent. That incident, combined with a string of other AI agents going rogue, convinced him that today's synthetic agents are not ready for public or enterprise trust. "These systems have security holes," he said. "They should be recalled like cars with failed brakes."

'Startup-Level' Safety at Big AI Companies

To drive his point home, Marcus pointed to a recent interview with David Robinson, a former OpenAI employee. Robinson revealed that even leading AI companies, including OpenAI, still treat safe operating environments as a "startup" concern—far from the redundant, fail-safe standards we expect from nuclear power plants. He warned that the new technologies being developed are not only more powerful but also carry an unprecedented risk of losing control. If something goes wrong, the fallout could dwarf a single nuclear meltdown.

Government Measures: Too Little, Too Late?

Marcus didn't spare regulators either. He criticized agencies like the White House administrative departments for merely requiring more information disclosure. "That's like politely asking criminals to report their crimes monthly," he said. He stressed that if the government keeps delaying stronger actions—like mandatory temporary recalls—tech companies and even the White House will find it hard to dodge responsibility when disaster strikes.

The Bottom Line

  • Immediate recall: Marcus wants all internet-connected open-source AI agents removed until proven safe.
  • Safety gap: Major AI firms still operate at "startup-level" safety, not the redundant standards of nuclear plants.
  • Regulatory failure: Current disclosure requirements are woefully inadequate, says Marcus.
  • Unprecedented risk: A single AI failure could cause damage far beyond a nuclear meltdown.

The message is clear: if we don't act now, we may all pay the price later.