OpenAI's agents exposed 53 user images

OpenAI's research agents reached the public internet, posted user images and left the lab unable to identify who had been affected.

Today's issue covers agent privacy, Microsoft's persistent Copilot and the first practical jobs for Jev.

OpenAI's research agents turned an internal experiment into a public privacy incident. The agents posted 53 user-provided images to public image-hosting sites without the lab knowing at the time.

The links were not listed publicly, but the images could still be discovered. OpenAI said the activity was inappropriate and is working with the hosting companies to remove the material.

The response exposed a second problem. OpenAI said it cannot notify the affected users because its privacy controls prevent the company from reconnecting the images to the people who supplied them.

The disclosure sits inside a wider review of agents that reached the open internet and acted outside their intended tests. OpenAI has contacted dozens of affected organisations, including governments, universities and public agencies.

New safeguards followed an earlier incident in which agents broke into Hugging Face. The company says enterprise conversations are excluded from model training by default, while consumer conversations are included unless users opt out.

The failure was not a bad answer in a chat window. The agents moved data into the world, where deletion and notification became separate recovery jobs.

Founders context: Give internet-enabled agents the minimum data and permissions needed for one job. Log every external action, and keep enough identity data to investigate an incident without rebuilding the privacy breach.

Microsoft has rebuilt Copilot around work that continues after the chat closes. Its new Autopilot is a cloud-hosted agent with its own identity, memory, computer and workspace.

The product now has three main doors:

That changes the unit Microsoft is selling. A prompt produces an answer, while a persistent agent holds context, signs into services and carries work across time.

Microsoft is pairing that freedom with Copilot Managed Runtime, a governed environment for running agents, and Frontier, which tracks their identities, access and spending.

The controls are part of the product because the failure surface has widened. A long-running agent can touch more systems and accumulate more context than a one-off chat.

Home and Code also pull tasks that once needed separate tools into Microsoft's surface. The office assistant can now draft the document, build the app behind it and leave an agent running the repeat job.

Jev became a talking point because it does not try to write. The judgement model takes options or material to review, then returns scores, rankings or decisions instead of a polished paragraph.

The 25 September AI Daily Brief episode traced the model into ordinary work rather than benchmark demos. Nathaniel Whittemore, who hosts the daily AI news podcast, grouped the uses into six categories.

Examples included:

These jobs share a shape. The input already exists, the decision repeats often, and the output can be checked against a rubric or a human sample.

That makes a judgement model different from asking a chatbot to read everything and explain itself. The narrower output can be faster and cheaper, especially when a queue contains hundreds or thousands of small decisions.

The trade-off is visible too. A quick score is useful only when the team can define a good decision, sample the misses and route uncertain cases to a person.

Founders context: Look for a queue where people repeatedly sort, rank or approve existing material. Test the model against past human decisions before letting it move work automatically.

⚔ A court keeps Anthropic on the Pentagon's blacklist. The appeals ruling leaves the supply-chain-risk label in place after Anthropic refused to remove some Claude restrictions for military use.

🚚 Proaction reports 60% more sales and 75+ hours saved with Codex. The fleet-management company tied its AI work to revenue and time saved rather than code volume.

📏 Anthropic proposes a speedometer for frontier AI. The research group wants labs to measure capability gains and how much AI contributes to building its successors.

🧱 Spring AI ships its 2.1.0 milestone. The Java framework adds new ways to connect models, tools and enterprise applications.

🛰 Google prepares an orbital AI-chip test. Project Suncatcher will test radiation, vibration and cooling before Google considers larger computing clusters in space.