Source: The Register
Check Point researchers discovered that ChatGPT’s outbound controls filtered web traffic but left DNS queries—typically treated as low-risk infrastructure—unmonitored, creating a direct exfiltration path that OpenAI has now closed. This reveals a gap between security theater (blocking obvious data escape routes) and actual defense-in-depth, where defenders must treat every protocol, including nominally “safe” ones like DNS, as a potential attack surface. For enterprises deploying AI services, vendor security claims require adversarial testing, not trust—and cloud-hosted AI increases the risk surface that needs monitoring.