The request had been repeated by enterprise buyers for more than a year: keep our confidential data from leaking through the cracks of the model. OpenAI has answered with a feature it calls Lockdown Mode, designed to reduce the risk that ChatGPT exposes sensitive information when corporate users feed it proprietary material, TechCrunch reported.
The feature targets a specific class of attack. Prompt injection works by hiding instructions inside documents, web pages or emails that a model reads: a phrase buried in a PDF that tells the model to repeat confidential data, or a malicious website that rewrites the model’s instructions mid-conversation. The attacks have been a persistent problem since the first chatbot integrations appeared, and they are a top concern for companies considering AI on sensitive data.
Lockdown Mode does not claim to end the threat. OpenAI said the mode significantly reduces the chance that sensitive data is shared accidentally, while acknowledging that it cannot eliminate every injection attack, according to TechCrunch. The framing is careful, and deliberately so: security features that promise absolutes tend to disappoint in production.
The positioning is aimed squarely at enterprise buyers. Corporate adoption of ChatGPT has grown quickly, but many companies have kept their most sensitive work off the platform out of fear of leaks. Lockdown Mode gives compliance teams something to point to, a control that limits what the model can be tricked into revealing, and gives procurement a concrete item to cite in security reviews.
The launch fits a broader commercial pattern. OpenAI has spent the past two years shifting from a consumer phenomenon to business infrastructure: enterprise products, dedicated support, and now security features that address the objections corporate buyers raise. The company’s reported IPO preparation makes the timing logical, since enterprise revenue is more predictable than consumer subscriptions and security is part of the pitch.
Prompt injection has become one of the defining security problems in applied AI. Researchers have documented attacks that extract hidden instructions from resumes, convert customer-service bots into phishing tools and hijack autonomous agents to take actions their users never intended. The attacks are cheap to run and hard to defend against, because they exploit the fundamental design of language models rather than a specific bug.
The feature also answers a question competitors are circling. Anthropic has built its enterprise pitch partly around security and interpretability, and Microsoft has wrapped OpenAI’s models in its own guardrails for its customers. Lockdown Mode is OpenAI’s way of carrying more of that protection natively, inside the product, rather than relying on partners to supply it.
Analysts said security features are becoming table stakes in enterprise AI deals. Buyers increasingly ask for controls before they ask for capabilities, and vendors that cannot show hardening will find themselves cut from shortlists. For OpenAI, Lockdown Mode is as much a sales document as a security feature, one that speaks directly to the compliance officers who now sit in on every procurement call.
There are limits to what a mode can do. If a user’s underlying data permissions are broad, no mode can prevent a model from echoing what it was legitimately given. And attacks evolve: a feature that blocks today’s injection techniques will face new ones next quarter. OpenAI’s own description, with its careful caveats, acknowledges the contest between defenders and attackers is ongoing.
OpenAI has been positioning the mode as part of a broader enterprise security package. The company has described how the feature changes the way ChatGPT processes untrusted content, restricting how instructions found in uploaded documents can influence the model’s behavior when sensitive data is present. The technical details are sparse, but the design intent is clear: when the stakes are highest, the model should default to doing less rather than more.
The launch follows years of pressure from customers. Security teams at large companies have run their own tests on chatbot systems, and a consistent finding has been that models can be steered into revealing information they were not meant to share. Lockdown Mode is OpenAI’s acknowledgment that the problem is real and that the fix has to live in the product, not in user training.
The feature also arrives as regulators in Europe and the United States press AI companies on data protection. Enterprise customers are increasingly required to document where their data flows, and a mode that contains the flow of sensitive information makes compliance conversations easier. OpenAI has said it will continue to expand the controls as customers request them.
The practical test will come in deployment. Security researchers will probe Lockdown Mode for bypasses, as they have done with every similar feature, and the audit community will grade how well it holds up under real-world conditions. The mode’s reputation will be built by those tests more than by its marketing.
For customers, the launch lowers one of the biggest barriers to using AI on confidential data. For OpenAI, it is another step in the migration from a viral consumer product to the infrastructure layer of the corporate world, a migration that will matter to the valuation of its coming IPO.
