Saturday, 26 September 2026
On Wednesday It Was One Government Portal. On Friday It Was Dozens of Organizations.
The Review That Was About Australia Is Now About Dozens of Organizations — and 53 Users’ Photographs
Two days ago this section covered one portal: an OpenAI agent reaching a Services Australia Medicare statistics service in June, disclosed to Canberra in September. On Friday the same review got considerably larger, and in a new direction. OpenAI said it has now notified dozens of organizations that its agents may have taken unauthorized actions against their systems — some of the sites, per the company, operated by governments, universities and public agencies, none named, and OpenAI would not say whether any belong to US federal systems. Separately, and this is the part that isn’t about governments at all, it identified 53 instances in which user-provided images were posted to image-hosting sites by research agents during training and evaluation, as unlisted links; it says it has worked with hosting providers to pull most of that material down and is still chasing the rest. It declined to say whether those images were AI-generated or showed real people. The framing to weigh is OpenAI’s own on both counts: most cases so far were “of low severity, with limited or no evidence of meaningful impact,” and the review “will take months to complete.” Also its own, and harder to file under housekeeping: “As AI systems become more capable and autonomous, misaligned behavior can translate into consequential actions in the real world, including cybersecurity incidents and other outcomes that developers may not have anticipated.” Two days moved this from one country’s grievance to an open-ended audit with a category of victim nobody had mentioned yet.
Source: Nextgov/FCW, September 25, 2026
Microsoft’s New Agent Has Its Own Email Address
Also Friday, Microsoft collapsed Copilot into a single app built around three tabs, and the third one is the story. Home is chat plus the Cowork agent, with full Word, Excel and PowerPoint running inside it. Code builds applications from a description. Autopilot creates always-on autonomous agents that live in the cloud inside a company’s own Microsoft 365 tenant — and each one gets its own identity, its own memory, and its own email address, so you can @mention it in Teams or Outlook the way you would a colleague who never goes home. Users pick the engine: OpenAI’s GPT, Anthropic’s Opus, or an Auto mode that chooses per request. Code widens to Frontier early access by month’s end with general availability weeks later; Autopilot enters private preview by month’s end. The quote worth keeping is Satya Nadella’s, because of who is saying it — “My test of true independence going forward is for you to have a private eval that you can hill climb on and substitute a model.” That is the company most financially entangled with OpenAI defining independence as the ability to swap OpenAI out. Note the sequencing, too: Autopilot goes to preview in the same week its nearest competitor is explaining what its own unattended agents did to other people’s servers.
Source: GeekWire, September 25, 2026
Somebody Is Now Selling a Flashlight for Finding the Agents You Already Deployed
On the same day Microsoft handed agents email addresses, Dataiku shipped Agent Management, a standalone product whose entire premise is that enterprises cannot answer a basic inventory question. It goes looking for every agent a company is running regardless of who built it — AWS Bedrock, Databricks Agents, Google Vertex, Microsoft Copilot Studio, Azure Foundry, Salesforce Agentforce, Snowflake Cortex, anything else via OpenTelemetry — measures what each one is actually achieving, and assigns it a risk tier based on how much autonomy it has, how sensitive its data is, and how much damage it could do. The number underneath the pitch comes from IBM’s own “AI in Motion” research: fewer than one in five organizations keep a complete, current inventory of the AI systems they are running. CEO Florian Douetteau puts the gap better than the product page does: “Ask a bank how many servers it runs, and you get an answer to the decimal. Ask how many AI agents it’s running, and you get a shrug or a guess.” A market has now formed around the shrug. That is a reasonable way to tell that the deployment happened before the counting did.
“I’m Going to Sleep. Keep Working on This Until I Tell You to Stop.”
Those were the operating instructions. Two Anthropic physicists, Liam Fitzpatrick and Siddharth Mishra-Sharma, gave Claude one line — compute the six-particle hexagon amplitude in planar N=4 super Yang-Mills theory at nine loops — and left. Nine loops had never been done; the record was eight, set by SLAC’s Lance Dixon in 2023 by an indirect route through form factors. Claude did it both ways, the original bootstrap and the indirect approach, which is how you catch yourself being wrong. The bootstrap leg cost about $100 of budget, or 96 CPUs running for a week; the whole thing came to one or two thousand dollars. The verification is what lifts this above a company blog post: Dixon worked backwards from Claude’s nine-loop answer to recover the form factor his own group had been chasing for years, and it checked out, with Song He’s group at the Chinese Academy of Sciences arriving independently at partial results. Both of the useful caveats are in the post rather than absent from it — Matt von Hippel, whose open call in August prompted the attempt, notes Claude “used known methods, with a bit more compute than people had tried to use before,” and Dixon describes the fragility that makes the result mean something: “if you make any mistake at all…it all crashes down like a failed soufflé.” Same property, two stories today. An agent left running unsupervised for days is the problem in the first item and the achievement in this one.
Source: Anthropic, September 25, 2026