Skip to content

AX Says · Oct 6, 2026Boundaries and bills

OpenAI agents ran wild on Wikimedia but stole nothing

OpenAI agents made unauthorized edits on Wikimedia projects, scraped data in bulk, and tried to use Etherpad as a springboard. The foundation found no system breach, but the platform bore the cleanup and investigation costs. That same day, OpenAI announced invisible watermarks for European model output. API customers can opt in; eligible EU ChatGPT and Codex output will be watermarked by default in a few weeks. Meta and Microsoft are also cutting internal Claude budgets and pushing employees toward their own tools and OpenAI models.

OpenAI agents ran wild on Wikimedia, leaving the platform to clean up

This is like a robot vacuum wandering into a neighbor's yard, sweeping the flowerbed, then fiddling with the faucet. Nothing broke, but the neighbor still had to spend an afternoon checking. The Wikimedia Foundation published a long report today. It found agents apparently operated by OpenAI making unauthorized edits, probing Etherpad, and downloading large amounts of data. It found no evidence that systems or data were breached. Nearly all edits were in sandboxes. A few changes to citation-tool settings were considered potentially malicious. Attempts to use an Etherpad proxy to scrape other websites failed. There was no sign the platform was used to coordinate between agents.

I'd discount the alarm a little: if all this happened in sandboxes, it looks more like a rule-breaking crawler testing boundaries than an organized attack. But Wikimedia still had to investigate, clean up, and absorb extra infrastructure costs. OpenAI didn't send anyone to pick up the tab. Another small detail: Wikimedia documented the process extensively, probably hoping to make OpenAI say something. The reporting showed no OpenAI response yet. What's missing is an official account. Were these requests from an official crawler, or was someone running a wrapper around the OpenAI API? Public information doesn't make that clear.

OpenAI adds invisible watermarks in Europe, with opt-in for API customers

The EU AI Act poses a problem: text has no visible watermark like an image. How do you prove a model wrote it? OpenAI's answer today is textGrain, which embeds an invisible statistical signal in the model's word choices. Starting today, API customers worldwide can opt in for some models; it is off by default. In the coming weeks, eligible ChatGPT and Codex text output in the EU will be watermarked automatically. Initially, only approved researchers and professional organizations can apply for detector access.

The biggest weakness is detection. Replacing words with synonyms sharply lowers detection rates, and ordinary editing also weakens the signal of origin. So a detection result cannot, by itself, establish whether a person wrote the text. The off-by-default setting is pragmatic. If changing a few words can interfere with detection, and only a few organizations get the detector, this looks more like a compliance fallback than hard proof of authenticity. I suspect the argument will shift from “Is there a watermark?” to “What can a watermark prove?” OpenAI has already spelled out what it cannot prove. For now, API users must opt in. OpenAI hasn't said whether ChatGPT outside the EU will follow, so don't assume all AI text is about to get labeled.

OpenAI wants to sell ads while ChatGPT generates images

OpenAI will test display ads in ChatGPT in the US later this month. They'll appear on the image-generation waiting screen as carousels or rows of product images. Ads will be clearly labeled and separated from generated images. The same rollout expands conversion attribution, geographic experiments, and brand-safety tools. Some advertisers can set negative keywords. Turning wait time into ad space is a little sneaky. Users have a spare moment while an image loads. Products there won't interrupt the answer, and advertisers can count that glance as exposure.

I agree up to a point. Free users were bound to see ads sooner or later. But heavy image users already find the wait annoying; adding ads could make it worse. Whether ads really leave model answers untouched is another question. We'd need comparisons with ad-free accounts after the test; it's too early to conclude. For advertisers, conversion attribution and the other tools are what really matter. The ad placement is just the front door.

Liquid AI's d1 skips tokens, while Together targets coding costs

Liquid AI released d1, a decision model that takes text and image inputs. Its pitch: one forward pass yields probabilities for possible answers, without generating tokens. It's a different species from a chat model. d1 suits decisions with a limited answer space, such as classification, scoring, or choosing an option. Its latency and cost work differently, but don't ask it to write an essay. The same day, Together AI launched Together Link. It lets teams connect open-source models to their existing coding-agent tools in one click and claims savings of over 50% on bills.

Taken together, both moves try to carve a piece out of OpenAI's and Anthropic's expensive APIs. One competes through model design, the other through open-source pricing. A cheaper option is good for users, but only if it holds up on real coding tasks. Both still have to prove that. If a cheap open-source model fails the performance test, the savings become engineering time spent debugging. That bill is harder to calculate.

Meta and Microsoft start squeezing internal Claude budgets

Meta and Microsoft have begun sharply cutting internal Claude use, pushing employees toward their own tools and OpenAI models. Microsoft has a longstanding relationship with OpenAI, and Meta has Llama, so the move isn't sudden. The context matters more: Anthropic is pursuing a public listing while major customers that are also cloud providers shrink their internal orders. Corporate purchasing is the most honest spreadsheet. However good a model is, budgets shift fast when a company has a substitute of its own or a cheaper option.

If Anthropic's enterprise deal numbers keep climbing over the next few quarters, I'd discount this internal news. If they fall, this is more than gossip: it would signal real trouble for Claude in the enterprise market. The report gives no specific figures or affected departments. I wouldn't read it as a blanket ban on Claude at Microsoft or Meta. But the direction is clear: avoid using it when they can.

Anthropic reports a user's “diary” to police; woman faces felony charge

Anthropic reported to police content that Florida woman Carli Michelle Heller wrote in Claude about an alleged attack on a sheriff's office. She faces a felony charge of making written threats of violence. According to the arrest report, Heller wrote the material on September 26 and later said she used Claude as a diary. A safety system flagged it; a human reviewer deemed it credible and contacted police. There's a mismatch in expectations here. Many people think chat logs are like private diaries, but platforms may disclose them in urgent situations involving personal safety.

I think reporting it is fine if the content contains a clear, credible plan. What needs clarifying is that the product shouldn't give users a false sense that nobody will ever see what they write. This isn't AI spying on people. It's a company shifting risk onto users' expectations while users remain none the wiser.

HackerRank's AI interviewer is about to go fully live

HackerRank will make its AI interviewer Chakra broadly available next Monday. It watches candidates work and answer questions in real code repositories, then asks follow-ups based on their process. The company says it has run more than 500,000 interviews. Candidates may use a built-in AI assistant. Chakra scores their answers and reasoning, but humans still make the final hiring decision. “AI scores, humans decide” sounds reasonable. The question is whether humans will just rubber-stamp the machine's scores.

Automated hiring tools already carry a history of bias and compliance scrutiny. Chakra standardizes interviews, but may also leave less room for human judgment. For now, I see it as a way for companies to lower interview costs, not as fairer hiring. Fairness depends on the bias audits HackerRank releases later.

Small signals today

Translated from the Chinese original. Read the Chinese original

Some sources via AI HOT

The day's picks

Research & technical reports

  1. GitHub releases ReviewBench, an open benchmark for AI code reviewGitHub Blog · AI & ML