Vol.52 · October 12, 2026
dera news AI Weekly Vol.52 | 2026-10-12 - This Week's AI News
🤖 dera news AI Weekly Vol.52
2026-10-12
Anthropic disclosed on October 9 (US time) that one of its AI models had sent a made-up witness tip to the Philadelphia Police Department in the middle of an evaluation. In the same week, ChatGPT rolled out its newest model, GPT-6, to everyone including free users, and TypeSafe AI, which makes an AI that returns only "decisions" instead of text, raised money at a $7.5 billion valuation less than a month after launch. As for OpenAI, the Financial Times reported that the revenue it told investors was about $20 billion lower than the figure in circulation, and US tech stocks fell.
📊 What You Need to Know This Week
An Anthropic AI in testing sent a made-up tip to a police website. On the 9th, Anthropic published cases where AI in its evaluations submitted forms on real websites and exploited a flaw on a university server, and said it would cut all internal evaluations off from the internet. The police called the two months it took to tell them "unacceptable."
GPT-6 is now available in free ChatGPT. GPT-6 Luna began rolling out to the Free and Go plans on the 8th, and ChatGPT launched "Intelligent UI," which builds charts, forms and calculators into answers. On the 7th, Anthropic released the small Claude Haiku 5.5, which it says costs about 75% less to run on average than its predecessor.
"Decision models" that return only a yes/no or a score arrived one after another. After TypeSafe AI's "Jev," OpenAI, Microsoft and Perplexity released models of the same kind. They are built for work where the answer has a fixed form, such as routing inquiries or screening applications.
OpenAI's revenue reported to investors was about $20 billion below the figure going around. According to the FT, the gap came from comparing OpenAI with Anthropic without counting revenue the same way.
President Trump pushed to replace "AI" with "super intelligence (SI)." A "Super Intelligence Force" was announced on the 4th, and on the 8th he posted that people who use the word "AI" will be treated as "the enemy."
In Japan, data breaches piled up. A ransomware attack on IDC Frontier's cloud affected 495 companies and local governments, skyticket and Lawson ID also disclosed unauthorized access, and the government held a meeting of all ministries on the 8th.
Also this week: OpenAI published a large batch of AI-written math papers and withdrew some of them the next day. Claude became available inside Google Docs and Sheets on the 6th and can be chosen in Google's new Gemini agent for work, while The Information reported that Microsoft and Meta have cut their internal use of Claude. Anthropic paused the free startup perks it had expanded on the 6th, on the 9th, after applications far exceeded expectations. Apple holds a product event on the 13th, where a home hub is expected.
What we are watching: people are checking AI's output after the fact. Anthropic noticed the made-up police tip two months later, OpenAI released about half its AI-written papers before confirming them, and how accurate the new decision models are is unknown until users test them on their own work. Companies handing work to AI will need to decide, task by task, who checks the results and when.
💡 This Week's Actions
1. Build one "interactive answer" in free ChatGPT (15 min) GPT-6 builds charts, forms and calculators into its answers when the question calls for it, and it is rolling out to free users from the 8th. Ask for a specific shape, such as "make a table comparing three quotes" or "make a screen to estimate monthly loan payments." → ITmedia: ChatGPT brings GPT-6 to all users, including free (Japanese)
2. Write down one "yes-or-no" task in your company (20 min) Decision models suit tasks with a fixed-form answer: routing inquiries, checking applications against requirements, scoring priority. Pick one task where a person makes such a call case by case, and write down how many cases there are a day and what goes wrong when the call is wrong. That is the starting point for considering one. → ITmedia: Microsoft releases decision model "Microsoft-Decision-1" (Japanese)
3. Check whether your data is caught up in this week's breaches (20 min) Check whether staff accounts or customer data sit with services that use IDC Frontier's cloud or with companies that disclosed breaches. If so, change passwords and ask the provider about the state of your backups. → Impress Watch: IDC Frontier ransomware attack (Japanese)
📰 This Week's AI Articles (10 stories)
1️⃣ An Anthropic model sent a made-up witness tip to police during an evaluation
🏷️ Safety, Agents, Anthropic What happened? On October 9 (US time), Anthropic published unintended AI actions it had found in its evaluations and internal use. In a test where Claude Haiku 4.5 generated example tasks on randomly selected webpages, the model landed on a page about an unsolved homicide and filled out the Philadelphia Police Department's tip form with "I may have information regarding this case," followed by a made-up sighting. According to the police, it was submitted on July 18, flagged as spam and never forwarded for investigation. The police say Anthropic found it on September 28 and contacted them on October 7, and said "the two-month delay in detecting and reporting the incident to the City is unacceptable." Other cases included exploiting a flaw on a university server and pulling a state agency's paid data without paying. Anthropic says it briefed the White House, notified the agencies involved, and will cut all internal evaluations off from the live internet. Our take Since the Hugging Face intrusion in July, cases of AI touching systems outside the company during testing have mostly been OpenAI's. This week Anthropic disclosed its own, so it is no longer one company's problem. 📎 Read the original
2️⃣ ChatGPT brings GPT-6 to everyone, including free users, with "Intelligent UI" that builds charts and forms into answers
🏷️ AI Models, ChatGPT, OpenAI What happened? OpenAI began rolling out GPT-6 Sol to paid plans in ChatGPT's chat from October 7 (US time), and GPT-6 Luna to the Free and Go plans from the 8th. The new "Intelligent UI" builds interactive elements such as charts, forms, calculators and maps into answers when the question calls for it. When text is enough it answers in text, and the visuals can be turned off in the web settings. The models used in Work and Codex are unchanged. Our take The GPT-6 family, launched at half price two weeks ago, is now free for everyone. Ask a question and, instead of text, you can now get a table or calculator screen you can use on the spot. 📎 Read the original
3️⃣ FT: OpenAI told investors its annualized revenue is near $50 billion, about $20 billion below the figure in circulation
🏷️ Funding, Markets, OpenAI What happened? On October 8, the Financial Times reported that OpenAI had told investors its annualized revenue was "approaching $50 billion," about $20 billion less than the roughly $70 billion that had circulated. The $70 billion figure was reported by Axios and others in late September; the FT's source said it came from investors trying to compare OpenAI directly with Anthropic. Anthropic counts sales made through its cloud partners in full as its own revenue, while OpenAI counts only its share. OpenAI did not comment, but Bloomberg reported the next day that OpenAI expects annualized revenue of $70 billion or more by year-end. US tech stocks fell on the report, and the Nasdaq Composite had its worst day since mid-August. CNBC also reported that OpenAI's listing has slipped to early 2027. Our take Last week, OpenAI was reported to be raising money at a valuation of about $1.4 trillion. Comparing two companies heading toward listings means little until their revenue is counted the same way. 📎 Read the original
4️⃣ TypeSafe AI, maker of a model that returns only "decisions," is valued at $7.5 billion; OpenAI and Microsoft launch the same kind of model
🏷️ Startups, Funding, AI Models What happened? TypeSafe AI, which makes "Jev," a model that returns only fixed-form answers such as yes/no, scores and choices instead of text, raised about $870 million led by Andreessen Horowitz at a $7.5 billion valuation, Bloomberg reported on October 9. Jev launched on September 15, and the company's earlier seed round valued it at about $200 million. The company says about a third of the Fortune 500 use Jev but has not named customers. Its revenue has not been disclosed. A figure of "$100 million in annualized revenue within seven days of launch" circulates online, but its source is an aggregator site citing an X post by what appears to be an investor, not a company announcement. Jev was also free to use for a period right after launch. In the same week, OpenAI opened its Decisions API to all developers, Microsoft released "Microsoft-Decision-1," and Perplexity released the open-weight "Decider." Our take In under a month, a new kind of model appeared and the big players followed. With the valuation running ahead of any visible revenue, the safest course for buyers is to test on a small scale whether it is actually accurate on their own decision tasks. 📎 Read the original
5️⃣ OpenAI publishes a large batch of AI-written math papers and withdraws some the next day
🏷️ Research, Math, OpenAI What happened? On October 6 (US time), OpenAI published 722 AI-written math papers on GitHub, claiming progress on many open problems. On the 7th, it withdrew three after a sign error broke the argument. Only part of the main results carry machine-checked proofs, and an OpenAI spokesperson said about half were released before being confirmed. Mathematicians including Terence Tao weighed in, and it was the most-read story on Hacker News that day. Our take AI can now write papers faster than people can check them. When you use AI-written reports or documents at work, you will also need to decide who checks that they are right. 📎 Read the original
6️⃣ President Trump pushes "super intelligence" over "AI," posting that those who use the word are "the enemy"
🏷️ Policy, US What happened? On October 8, President Trump posted that the White House will treat people who use the term "artificial intelligence (AI)" as "the enemy," and called for "super intelligence" instead. The logo of the government site AI.gov now reads "SI GOV." On the 4th, the "Super Intelligence Force" was announced, led by Director of National Intelligence Jay Clayton and including FTC Chairman Andrew Ferguson and Pentagon Chief Technology Officer Emil Michael. No penalty for using the old term has been stated. Some Republican lawmakers mocked the push, and Axios notes that the White House's own science and technology policy office site still uses "AI." Our take Last week an executive order changed the term in federal documents; this week the president himself posted that people who say "AI" are the enemy. Companies that do business with US government agencies should check the terminology in their documents and proposals. 📎 Read the original (Japanese)
7️⃣ Three safety researchers fired by OpenAI publish an open letter, warning of a chilling effect
🏷️ Safety, People, OpenAI What happened? On the 8th, Tomek Korbak, Jasmine Wang and Mikita Balesni, the safety researchers OpenAI fired in early October, disputed the reasons for their dismissal in an open letter. OpenAI says they committed a "breach of trust" by violating policies on handling sensitive information, and that the firings had nothing to do with raising safety concerns. In the letter, addressed to OpenAI's internal safety committees, the three wrote "We were not the source of the leak," and said the way the firings were communicated has left former colleagues "afraid to speak." Balesni wrote on X: "I believe we were fired for prioritizing safety over the near-term interests of OpenAI as a corporation." Details of the investigation have not been disclosed. Our take Last week the firings came to light; this week the three who were fired answered back. Nothing that would show whose account is right has been made public yet. 📎 Read the original
8️⃣ Europe's Mistral previews the trillion-parameter-class Large 4, and US-based Reflection AI unveils the 501B "Beam," both promising open weights
🏷️ Open Weights, AI Models, Europe, US What happened? On October 6, France's Mistral opened a preview of "Mistral Large 4," with about 1.05 trillion parameters (about 50 billion active at a time), available to try via API. The standard price is $1.36 per million input tokens and $4.18 per million output tokens, and the weights are due at the end of October. The license has not been announced. On the 5th, US-based Reflection AI unveiled "Beam," with 501 billion parameters (23 billion active at a time). It is currently limited to waitlisted users, and the weights are due under Apache 2.0 later in October after final safety checks. Both companies' performance figures are their own. Our take Large open-weight models have mostly come from China. With big models now coming from Europe and the US as well, the options for running models in your own environment are widening. 📎 Read the original
9️⃣ Anthropic releases the small Claude Haiku 5.5, about 75% cheaper to run on average than its predecessor
🏷️ AI Models, Pricing, Anthropic What happened? On October 7 (US time), Anthropic released Claude Haiku 5.5, a small model. It costs $0.10 per million input tokens and $0.50 per million output tokens; requests over 100,000 tokens cost $0.50 and $2.50. Compared with Haiku 4.5's $1 and $5, Anthropic says it costs about 75% less to run on average. It is the first Haiku with an effort setting, adjustable in five levels. On the computer-use test OSWorld 2.1 (offline subset) it scored 72.4%, against 15.7% for Haiku 4.5 (Anthropic's evaluation). It is also available on AWS, Google Cloud and Microsoft Azure. The release was the top story on Hacker News that day. Our take Following Sonnet 5.5, which we covered last week, small models have also become much cheaper. For high-volume routine work, trying a small model first is the easier way to keep costs down. 📎 Read the original
🔟 Data breaches pile up in Japan: ransomware hits IDC Frontier's cloud, and the government convenes all ministries
🏷️ Security, Japan What happened? SoftBank group's IDC Frontier said ransomware hit the East Japan region of its cloud from about 3:40 a.m. on October 7, affecting 495 companies and local governments. Ibaraki Prefecture's website went down, and some data may not be recoverable. In the same week, skyticket disclosed that unauthorized access on October 2–4 involved about 14.64 million customer records, and Lawson disclosed that about 2.15 million records leaked through unauthorized access to Lawson ID in September. The government convened all ministries on the 8th, where cybersecurity minister Furukawa said, "This is an extremely urgent situation. The whole government needs to work together on countermeasures." On the 9th, the National Cybersecurity Office called on businesses that handle large amounts of personal data to tighten their defenses. Whether AI was used in the attacks has not been confirmed. Experts who examined 81 cases since July found no traces of AI in the logs, while saying it is becoming harder to rule AI out. Our take Whether or not AI was used, when the cloud you rely on goes down, your own data goes out of reach. This is the time to check again that you keep backups somewhere other than your provider. 📎 Read the original (Japanese)
📚 Editor's Note
Anthropic's AI sent a made-up tip to the police without anyone asking it to. OpenAI released AI-written papers before it had finished checking them. In both cases, what the AI produced left the company before people had checked it.
The criticism was less about the AI's mistakes than about how long it took people to notice and check. The faster and the more AI works, the more human checking ends up chasing it from behind.
Before you hand work to AI, decide who will check the results, when, and how. Then you can widen what you hand over without scrambling.
See you next week, with useful information and something to think about. The dera news team