OpenAI is officially expanding its highly restrictive Lockdown Mode to all Free, Go, Plus, Pro, and Business accounts. The move directly targets a growing wave of prompt injection attacks, where hackers lace external webpages with hidden instructions to hijack the language model and secretly steal personal data.
Originally launched as an exclusive Enterprise feature in February 2026, this security update physically blocks the chatbot from transmitting compromised information back to a malicious server.
Security analysts note that Lockdown Mode operates deterministically rather than relying on the AI to spot a trick. By strictly limiting live outbound network requests, it successfully cuts off data exfiltration, meaning even if an injection compromises the prompt, the stolen information cannot leave the controlled network.
Cybersecurity experts highlight that intentionally severing the transmission path is what makes this defense work so well. It effectively breaks the so-called Lethal Trifecta of language model vulnerabilities, which requires access to private data, exposure to untrusted content, and a clear transmission path to succeed.
Alongside this network restriction, OpenAI is rolling out a new Active sessions manager between June 4 and June 5. This tool gives users the ability to review and remotely log out of unrecognized devices, according to a recent Engadget report.
How Lockdown Mode Disables Deep Research and Agent Mode
The trade-off for this deterministic security is a significant reduction in chatbot capability. Enabling Lockdown Mode requires users to trade advanced functionality for safety.
Turning on the feature immediately disables premium tools including Deep Research, Agent Mode, live Canvas networking, and the ability to download files directly from URLs. Live web browsing is also restricted strictly to cached content.
This creates a hard choice for power users who rely on the AI to execute live web tasks, forcing them to balance strict data protection against full system functionality.