Known AI failures should become requirements before the next deployment. That is the call from AI Incident Database founder Sean McGregor and Andrew Burt, CEO of AI-testing provider Luminos.AI, in a joint October 28 essay. They warn that organizations keep carrying familiar control problems into new systems.
The authors draw a line from harmful chatbot responses to agents that execute commands. Their central distinction is that a bad statement can be ignored, while an executed action can immediately change systems. Historical cases support their argument; the essay does not present a new experiment or measured incident rate.
McGregor and Burt propose practical steps for businesses. Before deploying an agent, decision-makers should compare its intended use with past incidents and ask vendors which failure classes they have mitigated and how they know. Lessons from other organizations should become contractual requirements, test plans and acceptance criteria.
They also argue against waiting for entirely new legal tools. Negligence, product liability and computer-intrusion law are among the possible levers they identify. That is their policy position, not a finding that a particular harm has already been adjudicated or that these tools have made an AI system safe.
The essay accompanies the authors’ Inside the Incident video series. McGregor heads the Responsible AI Collaborative, while Burt’s Luminos.AI offers agent testing. Their appeal focuses on the people making procurement and deployment decisions: turn a known incident into a requirement that can be checked.