OpenAI announced Lockdown Mode, a new opt-in security feature designed to reduce the risk of prompt injection attacks in ChatGPT. When enabled, the mode disables live web browsing (leaving only cached content), blocks retrieval and display of web images (image generation still works), and turns off deep research and agent mode. OpenAI acknowledged the feature does not make ChatGPT immune — malicious instructions could still appear in cached content or uploaded files and affect responses — but said the goal is to shrink the attack surface for users and organizations handling sensitive data. The company described the mode as not intended for everyday users, positioning it instead for enterprise and security-conscious customers.
OpenAI dropped Lockdown Mode for ChatGPT, a toggle that kills live web browsing, web image retrieval, deep research, and agent mode all at once. The idea: if you're handling sensitive data, you can strip out the features most likely to pull in prompt injection attacks. There's a catch — cached pages and uploaded files can still carry malicious instructions, so it's not bulletproof. OpenAI says the mode isn't for everyone; it's aimed at orgs and people who need stricter controls.
Fills a clear coverage gap in the tech category (0 stories in 48h) with specific, checkable claims from a named source (TechCrunch); the prompt injection angle is culturally relevant to internet culture's ongoing AI discourse.
Prompt injection has been one of the most discussed vulnerabilities in large language models since ChatGPT went mainstream — researchers have demonstrated attacks that exfiltrate private data through poisoned webpages and documents. OpenAI's move to offer a dedicated mitigation mode signals the company is treating the threat as a product-level concern, not just a research footnote. It also sets a precedent: competitors like Google and Anthropic may face pressure to ship equivalent controls, especially as enterprise adoption accelerates.
prompt injection is the LLM security boogeyman everyone's been talking about for two years and OpenAI just made it a product feature. shipping a toggle that strips out browsing, agents, and deep research is a real concession — it's basically saying the most powerful parts of ChatGPT are also the most attackable. competitors will get asked about this now.
Public story text does not change until an admin approves it.
Looped stories are not disposable posts: receipts, claims, reader checks, and moderator decisions can change the approved version over time.