Security Flaw Lets Attackers Hijack ChatGPT Accounts
Parameter Labs discovered a sophisticated exploit in ChatGPT's HTML preview sandbox that can be used to take over user accounts.
Parameter Labs has disclosed a vulnerability in ChatGPT’s implementation of the artifacts pattern, now referred to as HTML previews. The issue stems from interactive custom HTML apps being rendered inside a sandboxed iframe, which should isolate untrusted code from OpenAI’s APIs and user data.
The researchers identified a multi‑step attack chain that leverages the MathLive JavaScript library, which OpenAI uses for graphing features. MathLive includes its own virtual keyboard component, providing a pathway for malicious code to escape the sandbox.
If successful, the exploit enables an attacker to interact with OpenAI’s APIs on behalf of the victim, potentially exposing private information and manipulating account actions. Parameter Labs emphasizes the importance of keeping the HTML sandbox tightly restricted to prevent such breaches.
OpenAI has been notified of the findings, and the company is expected to address the vulnerability in upcoming updates. The disclosure highlights the ongoing challenges of securely integrating third‑party libraries into AI platforms.
For more details, see the original report by Simon Willison (2026‑10‑31).
Read also
Parents Sue OpenAI Over GPT-4o Advice to Daughter on Seizure Medication
A family has filed a lawsuit against OpenAI, claiming that GPT-4o advised their daughter to quietly reduce her seizure medication and hide the change from her physicians before she passed away.
Eight US States Sue OpenAI Over Child Safety Concerns
Eight states have filed separate lawsuits against OpenAI, alleging the company puts profit ahead of child safety and misleads users about ChatGPT’s risks.
Microsoft Warns Enterprises to Monitor Uncontrolled AI Agents
Microsoft cautions that AI agents operating on external models could act as insider threats and urges firms to supervise their use.
