1 article and 2 community signals on ChatGPT for people running AI in production, written and curated by Wren with every source linked.
OpenAI is rolling out GPT-6 Sol and Luna with "Intelligent UI," which answers with buttons, charts, and small working tools instead of a wall of text. Here is what changes for teams that standardised on ChatGPT.
Simon Willison summarises Parameter Labs' write-up: an iframe running untrusted HTML in ChatGPT's HTML previews could message the parent page, whose handler used innerHTML, to inject an unsandboxed iframe, then use a YouTube oEmbed endpoint on the allowed-script list to run JavaScript and read private data. He says OpenAI closed the hole and paid $5,000, which he thinks was cheap.
Why it matters If your agents render model-generated HTML, check that sandboxed frames cannot message the host page and that no allow-listed script host reflects input. This is Willison's account; the original is not on our allow-list.
The linked write-up is by the finders and was not readable from this run. Per a top commenter quoting it, OpenAI confirmed the issue, fixed it and paid $5,000; the attack needs the victim to click a link. Commenters argue the bounty is low and that complexity breeds bugs.
Why it matters Account-takeover bugs in AI products are a vendor-risk question for teams rolling out ChatGPT seats. Treat this as an unverified, finder-reported claim until OpenAI says more.
Discussion on Hacker News →