AIDEFEND in Action AIA-0136 Validated Research

ChatGPT Lockdown: HTML Preview Popup Escape Crossed the Outbound Boundary

Check Point demonstrated a ChatGPT HTML-preview popup escape that bypassed Lockdown restrictions. A new nested iframe supplied an unmodified popup function, while permissive sandbox tokens let the new window escape restrictions. The reported Electron-based ChatGPT Classic allowed a no-gesture route; ordinary browsers generally required a click. Phishing and malware delivery had further user-action conditions. OpenAI reported a September 18 fix, and the researcher confirmed the technique no longer worked.

Guardrail BypassData ExfiltrationSocial EngineeringClient SecurityWeb Security
3 applicable AIDEFEND defenses

Threat Analysis

  • The enforcement gap was below the prompt: HTML-preview scripts overrode one popup function, but a new nested frame exposed another. Permissive iframe tokens let a popup escape the intended restriction.
  • Client conditions change impact: The tested Electron Classic client allowed a no-gesture route. Ordinary browsers generally required a click; Classic downloads required another click. Phishing still needs credential entry, and download is not execution.
  • The disclosure includes a repair: The researcher reports OpenAI’s September 18 removal of outer-frame popup permissions and confirms the original technique no longer worked. Parent-origin cookie or localStorage theft was not demonstrated.

Applicable AIDEFEND Defenses (3)

AID-H-003.010
Deployed AI Software Vulnerability Remediation Lifecycle
Very High
The provider must deploy the repaired outer-frame policy to every affected preview/client surface and independently verify effective runtime flags. The researcher reports removal of both permissive popup tokens and a failed post-fix reproduction. Verify normal previews still work and the nested-frame route cannot open an escaped window. Customers should use supported current clients; the source supplies no client build number to invent.
AID-H-006.002
Text, Markup & Structured Output Sanitization and Release Gate
Very High
For output channels that do not require active HTML, the provider or an independently built assistant can render generated code as escaped text at the actual HTML sink. Route every such channel through the encoder and reject unknown sinks. This prevents the script from becoming executable markup; it deliberately changes an active preview into a non-executing display. For a constrained output channel, remove active frames and remote-fetch surfaces before rendering, disable automatic previews and enforce a restrictive isolated-origin policy. Verify real browser egress independently, including nested frames, popup/navigation attempts and both client gesture modes. The reference sanitizer is a non-active rendering option; keeping arbitrary JavaScript requires the repaired provider sandbox and cannot be declared safe from a sanitizer-only test.
AID-R-004
Fleet Remediation Propagation & Technical Recurrence Prevention
Medium
The provider should retain a harmless regression corpus for nested-frame popup recovery, no-gesture Electron behavior, browser-click paths, spoofed navigation and download handling. Test both the repaired boundary and legitimate preview functions after client or renderer changes, with synthetic markers and no real credentials. Record the exact client/runtime build and independent network/window observations.

What Defenders Should Do Now

  • Provider teams should verify the repaired iframe and client policies across all preview surfaces. For channels that do not need active HTML, escape text and remove remote/active markup before the actual rendering sink.
  • Test child frames and every created window, not only the original page’s JavaScript hooks. Include independent egress observation and both browser and Electron interaction conditions in release regressions.
  • Users should keep supported clients current and treat preview update/sign-in prompts as untrusted. If a real incident includes entered credentials or an executed download, investigate that specific exposure separately; this disclosure itself does not establish such victims.

Conclusion

AIDEFEND  places the effective controls at software remediation, generated-output release and regression verification. The decisive boundary is the client’s actual rendering and window behavior, with provider-owned enforcement and explicit per-client limits.

Last updated: