ChatGPT now lets you virtually try on clothes using your own photos. OpenAI rolled out a virtual try-on feature inside ChatGPT that lets users upload a photo of themselves and see how clothing and accessories would look on them. A Favorites library lets users save products they like across sessions. The feature is available now and sits inside ChatGPT's existing shopping interface. It puts ChatGPT directly in the path of a purchase decision, not just the research phase. (TechCrunch)
Brief
Friday, October 2, 2026 · 4 min read
OpenAI Fires Three Safety Researchers After Data Leak
Good morning. It is a Friday in October 2026 and the week's defining move was not a new model, it was a grocery chain and a social club filing paperwork faster. Meanwhile OpenAI parted ways with three safety researchers, which is the kind of news that lands quietly on a Friday afternoon for a reason.
Today's reading time is 4 minutes.
INDUSTRY
AI labs have spent two years arguing that safety and commercial ambition can coexist inside the same org chart.
Driving the news: OpenAI cut ties with three safety researchers following an internal investigation that found they mishandled sensitive company information, according to a Wall Street Journal report picked up by TechCrunch. The company has not named the individuals or specified what information was involved. The departures come roughly a week after OpenAI's DevDay conference, where the company was publicly emphasizing its responsible deployment story alongside a wave of new agent features.
- Three researchers gone, all from the safety function, per the WSJ report.
- The stated reason is mishandling of sensitive internal information, not performance.
Zoom in: OpenAI has had a complicated relationship with its safety team for a while. Several high-profile safety researchers left voluntarily in 2024 and 2025, some publicly citing concerns about the pace of deployment relative to safety work. Firing three researchers, rather than watching them leave, is a different kind of event. It puts the company in the position of defending both the decision and the circumstances that led to it.
- The timing, one week after DevDay, means the story competes with a lot of positive product coverage.
- OpenAI's safety team has been a recurring subject of public scrutiny since the Ilya Sutskever era.
Why it matters: For anyone building on OpenAI's APIs, the safety team's health is a practical concern, not just a PR one. Safety researchers are often the people who catch model behaviors before they reach production. Fewer of them, under whatever circumstances, is a thinner buffer between a model quirk and a shipped product.
Bottom line: OpenAI can call it a misconduct case, but three safety researchers leaving involuntarily in the same week is not a footnote.
TechCrunch ↗Get this every weekday
Two minutes, 7am ET. No fluff.
Also happening
Suno added spoken voice generation to its AI music platform. Suno, best known for generating full songs from text prompts, launched a public beta that generates spoken voiceovers alongside background music in a single pass. Users can provide a script or a prompted description and get synchronized voice and music out. The feature is live on Suno's web and mobile platforms now. It is a direct move into territory occupied by dedicated text-to-speech tools, without requiring a separate workflow. (The Verge)
Shopify launched Canvas, a store builder you configure by chatting with an AI agent. Canvas lets Shopify merchants describe what they want and watch their storefront update in real time, driven by Shopify's Sidekick agent. It is a direct replacement for the drag-and-drop store editor for merchants who prefer to work in natural language. The feature was announced yesterday. It is the clearest sign yet that Shopify is betting its merchant base will adopt conversational interfaces faster than traditional no-code tools. (TechCrunch)
A patched vulnerability in ChatGPT's Mac app could have exposed sensitive user data to attackers. Wired reported on a now-fixed security flaw in the ChatGPT desktop app for macOS that could have allowed a malicious actor to access sensitive data handled by the app. The vulnerability has been patched, but it illustrates that AI applications are themselves attack surfaces, separate from any capabilities the models have. The timing is awkward given the broader conversation about AI agents being given more system access. It is a reminder that the app wrapping the model needs the same scrutiny as the model itself. (Wired)
The thread
AI moving into the purchase path
Two of today's stories put AI tools directly between a user and a buying decision. ChatGPT's virtual try-on feature sits inside a product search that ends with a purchase. Shopify's Canvas lets merchants build the storefront that receives those buyers, also through an AI interface. The research-to-checkout gap is getting shorter, and the tools controlling that gap are increasingly the same ones people use for everything else.
Try it yourself
Virtual Try-On
ChatGPT's new try-on feature requires uploading your own photo to one platform; Snippt's virtual try-on tool lets you do the same thing without a ChatGPT subscription.
On our radar
- Google's Guided Vision feature in Gemini Live uses your phone's camera to read fine print and describe surroundings aloud in real time, available now on compatible Android devices.
- Cohere released Embed 5 in two tiers, Pro for retrieval quality and Fast for low-latency query paths, targeting enterprise search and RAG pipelines.
- AWS open-sourced Strands Decider 2B, a decision model that returns choices and confidence scores in 115 milliseconds on an RTX 3090, under Apache 2.0.
- A new arXiv paper found that detailed profession-specific system prompts raise token costs without consistent accuracy gains across 503 tested agent profiles.
- Airbnb CEO Brian Chesky argued in a TechCrunch interview that AI agents need a dedicated operating system layer, and that Airbnb is actively making its platform agent-accessible.
- Albertsons is using ChatGPT Enterprise across its teams to speed up internal workflows, per an OpenAI case study published yesterday.
- Wired's Uncanny Valley podcast covered the voluntary AI safety accord signed by tech executives and the rise of consumer-facing AI agents, including a discussion of the US midterm ballot.
Get the brief in your inbox
Every weekday morning. Two minutes, no fluff.