OpenAI adds Lockdown Mode and risk labels to ChatGPT

OpenAI has introduced two new security features in ChatGPT — Lockdown Mode and Elevated Risk labels — designed to help organizations defend against prompt injection and AI-driven data exfiltration.

Prompt injection, in which hidden or crafted instructions trick a model into ignoring its intended guidelines, has become one of the most persistent attack vectors as companies connect large language models to internal tools, documents, and external websites. Data exfiltration attacks exploit that same mechanism: instead of merely manipulating model behavior, they push sensitive information — file contents, credentials, conversation history — out to attacker-controlled destinations.

Lockdown Mode gives administrators a stricter operating profile for ChatGPT deployments where the threat of manipulation is high. Elevated Risk labels, in turn, flag activity that suggests a higher likelihood of an attempted attack, giving security teams a signal to act on rather than forcing them to audit every interaction manually.

The release matters because enterprise adoption of AI assistants increasingly hinges on whether security teams can trust those systems with access to corporate data. Agentic workflows — models that browse, read documents, and call tools — widen the attack surface for injection and exfiltration, and buyers have been demanding finer-grained controls from vendors. OpenAI shipping these features signals that defensive tooling for LLMs is moving from research papers and startup products into the default feature set of a mainstream consumer and enterprise platform.

The announcement does not specify rollout timing or licensing details, so organizations will need to watch OpenAI's enterprise documentation for when the controls land in their admin consoles. Expect rivals such as Google and Microsoft to face similar pressure to match these defenses as prompt injection stays at the top of AI security threat models.