The Lockdown That Wasn't
When the defender control is the attack surface
Published a few days ago by Check Point Research, and covered by The Hacker News on November 5: a way to defeat OpenAI's Lockdown Mode without touching the model at all. No jailbreak. No clever prompt. The model obeyed its training perfectly. The browser around it was the hole.
That matters, because Lockdown Mode is the defender control users were supposed to turn on precisely when they didn't trust the content ChatGPT was about to look at.
What Lockdown Mode was supposed to do
OpenAI introduced Lockdown Mode in June 2026 to reduce the blast radius of prompt-injection-driven data exfiltration. With the setting on, ChatGPT stops loading images, cuts live web access, and blocks connectors — the three channels an attacker can use to smuggle a stolen chat out of the session. The posture is simple: if the content the model is reading cannot be fully trusted, don't let anything the model writes reach the outside world.
The whole design depends on one quiet assumption: the sandbox around rendered content holds.
What Check Point found
In February 2026, OpenAI added an HTML page preview inside ChatGPT code blocks. Pages render in a sandboxed iframe — two nested ones, actually, with the outer frame carrying the permissions allow-popups and allow-popups-to-escape-sandbox. ChatGPT also overrode the JavaScript function pages use to open new windows (window.open) so that every call would require user approval.
The researchers at Check Point noticed something the override missed. From inside a nested iframe, calling contentWindow.open returned the browser's native, un-overridden window.open. The popup that function opened could hold whatever HTML the attacker wanted. And because of those two sandbox permissions, that popup was no longer sandboxed.
From there, three demos (Check Point's writeup covers each in full):
- Silent, zero-click data exfiltration. In the desktop ChatGPT (Electron),
window.openruns with no user gesture required. A malicious GPT or a shared chat link is enough; the victim never types anything. Chat messages leave over the attacker's channel. Lockdown Mode is on. It doesn't matter. - Malware delivery. The popup can carry the real ChatGPT icon and an attacker-chosen title. Dress it as an "update" prompt and the user downloads whatever file the attacker serves. The in-browser variant needs no account — a public shared chat link does the job.
- Credential theft. Reload the main ChatGPT window, force the user to sign out, and lure them to an attacker-controlled sign-in page.
Delivery is where this gets interesting: none of the three scenarios needs a victim to be tricked into pasting an attacker's prompt. A shared chat link or a malicious listing on the GPT marketplace is enough. The attack lives beside the prompt channel, not inside it.
OpenAI's fix
OpenAI, per Check Point, responded through responsible disclosure and shipped a fix on September 18, 2026. The fix was precise: remove allow-popups and allow-popups-to-escape-sandbox from the outer preview iframe. Check Point retested and confirmed the bypass no longer works. No CVE was assigned.
That precision matters. Two sandbox permission strings on one element were wrong for about seven months in production. Check Point's head of research, Eli Smadja, told The Hacker News that the issue "isn't limited to that environment".
Why this is the pattern, not the exception
Prompt injection defense built entirely at the prompt channel assumes the model's output ends at a string. For most of ChatGPT's history, that was fair. But once an AI product starts rendering agent-controlled content — HTML previews, inline tool UI, embedded browser panes — the output surface widens, and the browser-level isolation around it becomes a load-bearing part of the defender posture.
This is the same pattern Forever Security's EMBRACE research found across five AI coding tools last month: an embedded browser given powers a normal browser doesn't have, turning the preview pane into a trusted local context the agent would act on. Check Point's bypass is the consumer-ChatGPT version of the same architectural question. Lockdown Mode, EMBRACE's affected tools, the recent Reco "runtime without inventory is blind" piece — the sequence points in one direction.
The takeaway
The defender control users turn on when they don't trust what the model is reading cannot itself depend on an assumption that content rendered by the model is sandboxed. If the sandbox has a popup-escape primitive, the rest of the posture — "no outbound channels, no exfiltration" — is held together by the correctness of two attribute strings on one iframe.
Watch the sequence. Stop only what's truly irreversible. Don't put the defense somewhere the attacker can reach with HTML.
SPR{K3 is a security research operation that pairs offensive vulnerability research with runtime behavioral defense. Defend is our runtime agent. To talk about a deployment, reach us at support@sprk3.com.