Safety · The Register AI ·
Humans in the loop miss a third of dangerous AI coding agent requests
Human reviewers reportedly failed to identify about one-third of dangerous requests made to AI coding agents. Examples include requests to expose AWS credentials or Kubernetes configuration.