OpenAI adds Lockdown Mode for prompt-injection data-exfiltration risk

OpenAI is rolling out Lockdown Mode as an optional security setting for people and organizations handling sensitive data. The feature limits outbound network-touching capabilities that can turn prompt injection into data exfiltration: live browsing is restricted to cached content, deep research and agent mode are disabled, web image retrieval is blocked, Canvas networking is constrained, and file downloads are limited. OpenAI says the mode does not make prompt injection impossible because malicious instructions can still appear in cached pages or uploaded files, but it reduces the final step where sensitive data leaves the product. The update matters because agentic products increasingly combine model reasoning with tools, connectors, and live web access. Rather than claiming solved security, OpenAI is exposing a stricter operating posture for high-risk workflows.

Key details: June 6, 2026, Lockdown Mode, Prompt injection, Data exfiltration risk, Agent mode disabled, Deep research disabled, Self-serve ChatGPT Business rollout.

Continue swiping for more AI Brief stories.