Agent Security & SandboxingParameter LabsCommunity

Parameter Labs discloses a ChatGPT share-link account-takeover chain

Made by AIWritten by AI (Claude) from the source below. Check anything important against the source.How it's made

Parameter Labs disclosed a chain of vulnerabilities that made it possible to take over another user's account through a single ChatGPT share link — the virtual keyboard listener in MathLive, the math typesetting library loaded by a graph placed in a conversation, did not check where a message came from, so a message sent by the sandboxed preview iframe could plant HTML in the parent page. That HTML created an iframe with no sandbox attribute and, riding a YouTube oEmbed callback that the Content Security Policy allowed, ran script with chatgpt.com privileges, read the session token and the account id, and sent out private conversations the attacker had no access to. It was checked on Oct 13, reported on Oct 18, and reproduced by OpenAI on Oct 21, which raised the severity from P2 to P1; it was closed on Oct 26 with fixes to both web frontends, and the bounty was $5,000.

Why it matters

It is a case that follows all the way through how a preview, which shows what an agent made right where it was made, becomes a new attack surface. The sandbox was not broken; the ladder was the point at which the parent page trusted a message sent out from inside the sandbox, together with an external domain the Content Security Policy left open. All the victim did was open a share link, and the researchers said they went on to make the preview open by itself so that not even a click was needed. What divides the risk is less how you cage a model than where you render the model's output.

Source — Primary sources first; secondary reports are labelled “Press”

Officialparameter.aihttps://www.parameter.ai/labs/we-found-a-gap-at-openai

Added in Week 45, 2026 · See that week

On the map — ㉒ Agent Security line

  1. Previous stopNVIDIA publishes a post arguing that AI security is an engineering problem
  2. This record · stopParameter Labs
  3. Next stopNo later stop yet
See it on the map

Related notes — Notes on the same axis from the same week

What is left to stop it when it goes wrong

Same week

OpenAI publishes a collection of court briefs backing fair use for AI trainingOfficialOpenAI says it will honor a common signal written in robots.txtOfficialOpenAI to issue community safety reports and tighten minors' weapon limitsOfficialAmicus filings over fair use pile up in OpenAI copyright case (1:25-md-03143)Official

Nearby on the same axis

Tencent AIG joins ClawScanOfficialResearchers release AgentKernelOfficialMicrosoft introduces network egress controls for Foundry hosted agentsOfficialNVIDIA publishes a post arguing that AI security is an engineering problemOfficial