OpenAI is shipping Lockdown Mode, a security setting that trades web access for protection against prompt injection attacks.
Prompt injection is the attack class where malicious text embedded in a webpage, document, or tool output hijacks an AI model's behavior, redirecting it to follow an attacker's instructions instead of the user's. OpenAI's answer is to eliminate the attack surface: Lockdown Mode disables live web browsing and web image retrieval. The result is a model that cannot be manipulated through content it fetches from the internet, because it no longer fetches it. No magic, just removal.
The timing matters. AI agents that browse, summarize, and act on web content have become a mainstream product feature, and prompt injection is the obvious weak point security researchers have flagged for years. A compromised agent with access to a user's tools or accounts is not a hypothetical; it is a demonstrated attack path. Shipping a named, opt-in mode signals that OpenAI considers the threat real enough to require a hard countermeasure, not just a blog post.
The honest read: this is a mode that solves the problem by unplugging the vulnerable parts. Useful for high-stakes deployments where auditability matters more than convenience, less useful for anyone who bought into the pitch of an AI that can browse the web on their behalf.