theaisafemode.com
AI Safe Mode

A Chatbot Security Mode Had a Gap Outside the AI

Colourful lines of code on a screen; illustrative
Illustrative photograph only, not an event image. Photo by Ilya Pavlov via Wikimedia Commons; CC0 1.0. See LICENSES.md.

Researchers disclosed a now-patched preview flaw that bypassed a chatbot’s extra security mode. The lesson is to check the whole app, not just its model.

An extra security setting did not cover every route out of a popular AI chat app. Researchers disclosed a flaw that let specially prepared preview content bypass that setting and create opportunities for data theft and deceptive popups. The issue has been patched.[1] [2]

The Hacker News reported the disclosure on Nov 5, followed by TechRadar on Nov 6. This is a report about a repaired weakness, not a warning that the demonstrated attack still works today.[1] [2]

What the researchers found

The weakness involved the way the app displayed web-page previews inside code blocks. Those previews were meant to run in an isolated area, with restrictions on connections and access to the rest of the application.[1] [2]

Researchers found that the handling of popup windows offered a route around those restrictions. They described possible data theft, phishing and malware-delivery scenarios, including deceptive content that appeared within a familiar application environment.[1] [2]

The reports say the provider added protections after the researchers disclosed the issue. They do not establish that the technique was used to compromise ordinary users in the wild.[1] [2]

Why it matters

This was not simply a chatbot choosing the wrong words. It concerned the software surrounding the model: previews, windows and the controls that separate content from the rest of an app.[1] [2]

For users, the practical lesson is to keep several questions separate. Is the app up to date? Is the content trusted? What information can this session reach? An extra security setting is worth using, but do not use its reassuring name as permission to paste every sensitive document into a chat.

There is also a useful distinction between a research demonstration and a confirmed compromise. Do not assume that reading this disclosure means your own account was attacked. Review actual account activity and follow the provider’s current guidance rather than installing a supposed repair from an unsolicited message.

Check a warning without following its instructions

If a shared conversation opens a window that asks you to update, sign in again or download something, pause before interacting with that window. Close the suspicious content and reach the service through your usual trusted route. Look for information about updates or account security there rather than using the link supplied inside the conversation.

Separate the appearance of a window from the action it requests. Familiar colours or an application icon are not the verification step. Ask whether you expected this request, what it would allow and whether you can complete the same check through the normal service interface. If you cannot establish that, do not provide credentials or run the proposed download.

For a work account, give the responsible support team the minimum relevant information: when the window appeared, what you were opening and what action you took, if any. Do not post a private chat or a screenshot containing confidential data in a public help forum. Preserve evidence in the channel your organisation uses for security concerns.

If nothing suspicious happened and you are simply reading the disclosure, keep your response proportionate. Follow official update guidance and review unnecessary connected services. Do not search for an unofficial test or attempt to reproduce the weakness in your own account. The useful outcome is a better verification habit and a narrower set of exposed information, not an experiment with a technique the reports say has been fixed.

How to stay safe

  • Use official updates. Update through the service or your normal trusted app-distribution route.
  • Distrust in-chat update requests. Do not download a repair because a shared conversation or popup asks you to.
  • Limit session access. Keep unnecessary accounts, private files and long-lived credentials out of the session.
  • Treat shared content cautiously. Open material only when you understand its source and purpose; avoid running unfamiliar previews for curiosity.
  • Check real evidence if concerned. Review sign-ins and connected services, and use official support for suspected unauthorised activity.

The short version

The flaw was patched. The durable lesson is that an AI app’s surrounding features need security too, and no single setting replaces careful access control.

Follow AI Safe Mode

Protect yourself with Privaku

See the post on YouTube Watch and follow AI Safe Mode on YouTube @AISafeMode and Instagram @aisafemode.

References

  1. The Hacker News, Nov 5, 2026. Report on a patched chatbot security-mode bypass. Read the report ↗
  2. TechRadar, Nov 6, 2026. Report on preview-window security research. Read the report ↗
Facts come only from the published reports above and are attributed to them. AI Safe Mode doesn't name companies, products or people involved in incidents. Spotted an error? Email [email protected] and it will be reviewed and corrected. See the corrections policy.