Security researchers have found a flaw that could allow attackers to bypass ChatGPT's Lockdown Mode and steal users' private chat data, including through zero-click attacks on desktop.
Security research · ChatGPT Lockdown Mode
A flaw could let attackers bypass ChatGPT Lockdown Mode and quietly steal private chat data
Researchers report a weakness in the HTML preview feature that enables silent data exfiltration, including a zero-click variant that affects desktop users.
How the reported attack path works
Simplified from the findings; full technical mechanics have not been disclosed.
Step 1
HTML preview renders
Rich content is displayed inside the chat experience.
→
Step 2
Silent exfiltration
Data leaves the chat with no visible sign to the user.
→
Step 3
Private chats exposed
Conversation content reaches an attacker.
Most concerning variant: zero-click. No action is needed beyond normal use, such as viewing content.
Why it matters
Lockdown Mode exists to limit risky behavior, so a bypass removes protection users reasonably expect.
The exposed data is among the most sensitive users share with AI assistants.
The finding gives security teams a concrete case for independent testing of protective modes.
Open questions
Full technical mechanics of the bypass are not specified.
The number of users potentially affected is unknown.
It is unclear whether a fix has been deployed.
Key takeaway
A security control was circumvented, and that exposed conversation content.
Exposure
Personal, professional, and confidential conversations.
Severity driver
Zero-click, silent, and reported on desktop.
Broader lesson
Every rich-content rendering path is a potential leak route.
Continue reading The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.
Already purchased? Sign in ✓ Signed in — this article isn’t included in your current plan.Unlocking the full article...