OpenAI's new Lockdown Mode restricts external capabilities in ChatGPT and agentic tools to reduce data exfiltration risks, but security analysts and CISOs are skeptical of its effectiveness. The mode limits web browsing, disables Deep Research and Agent Mode, and restricts file downloads, yet experts note it still allows side-channel exfiltration. Controversy surrounds OpenAI's own FAQ claiming 'prompt injection is not currently a major risk' while simultaneously building containment features for it. Analysts are divided on whether enterprises should rely on vendor-provided controls or their own network segmentation and Zero Trust approaches. The broader challenge is that multi-vendor AI environments will create a patchwork of incompatible lockdown controls, pushing organizations toward AI-specific governance frameworks analogous to network segmentation and least-privilege models.
