OpenAI Help: Lockdown Mode

The feature constrains outbound requests after an injection is already underway, meaning the attack surface it addresses is narrow: exfiltration specifically, not the injection itself. That distinction matters enormously for security teams evaluating whether this changes their threat model in any meaningful way.

OpenAI has been on an aggressive infrastructure and distribution push this month, from the AWS Marketplace partnership announced June 1st to the Michigan Stargate data center. Lockdown Mode fits that pattern: it's a feature that makes ChatGPT more palatable to enterprise procurement and governance workflows, which is exactly the audience the AWS deal is courting. The timing is not coincidental. Meanwhile, the arXiv work on eating disorder query failures from the same week is a reminder that OpenAI's safety investments remain uneven, hardening one attack vector while systematic alignment gaps in sensitive domains go unaddressed.

Watch whether Anthropic or Google DeepMind ship a comparable named feature within 90 days. If they do, this becomes a baseline expectation rather than a differentiator, and the real competition shifts to which platform can demonstrate measurable reduction in successful exfiltration attempts in third-party red-team audits.