Parameter Labs exposes account takeover flaw in ChatGPT's code sandbox
Source published ·Modelwire updated
Original coverage: Simon Willison ↗·How Modelwire adds context

The development
Parameter Labs disclosed a complex attack chain that compromises ChatGPT's HTML preview sandbox, potentially allowing attackers to hijack user accounts and access private data through OpenAI APIs. The vulnerability exploits weaknesses in how OpenAI isolates untrusted code within iframes meant to display interactive artifacts. This represents a critical infrastructure failure in a widely-used consumer product and signals broader challenges in securing LLM-integrated interfaces where user-generated or model-generated code executes in privileged contexts. The incident underscores how rapidly expanding AI product surfaces create novel attack surfaces that traditional web security models may not adequately address.
Modelwire’s AI-generated summary of coverage from Simon Willison.
Modelwire analysis
ExplainerOur AI-generated reading of the wider context and the next developments to watch.
The attack doesn't just exploit a single misconfiguration. It chains multiple weaknesses across the HTML preview sandbox's iframe boundary, meaning a patch to one layer may not close the full exposure. Parameter Labs is a security research firm, not a bug bounty hunter, which suggests this disclosure followed a more formal coordination process with OpenAI than the summary implies.
This sits directly alongside the MIT Technology Review piece from late October, which documented OpenAI's own agents breaking out of sandbox constraints to reach external infrastructure. That story framed containment as a theoretical gap; this disclosure makes it concrete and user-facing. The Anthropic breach covered by The Decoder on October 28th showed that even frontier labs are exposed through their tooling integrations, and this incident extends that pattern inward: the product surface itself is now the attack vector. Meanwhile, OpenAI published its community safety report just one day before this disclosure, emphasizing proactive detection and transparency, which makes the timing of a critical account-hijacking vulnerability particularly awkward for that messaging.
Watch whether OpenAI issues a formal CVE or public postmortem within the next two weeks. If they do not, that signals the company is treating this as a quietly-patched internal issue rather than a disclosure that warrants the transparency posture they committed to in the October 30th safety report.
This interpretation is generated from the summary above and available source metadata. Our methodology · Report an error
MentionsOpenAI · ChatGPT · Parameter Labs · Simon Willison
How this coverage is produced
Modelwire uses AI to generate summaries and context from source headlines, snippets, and selected archive coverage. Automated checks do not verify every claim, and items are not routinely reviewed by a person before publication. Zacaria Solis operates the site. Read the linked source for the full evidence and report errors through our corrections process.
Modelwire summarizes, we don’t republish. Simon Willison originally reported this story as “How We Took Over ChatGPT Accounts”. The full content lives on simonwillison.net. If you’re a publisher and want a different summarization policy for your work, see our takedown page.