AI News06 Oct 2026
None of the AI giants told New York they were insured against disaster
AI's biggest risks are usually debated at conferences, where nobody speaks under oath. Yesterday New York's city council put Anthropic, OpenAI, Google and Meta under oath, after threatening them with subpoenas. A former Anthropic researcher told the council that humans losing control of AI is more likely than not. The council wants outside checks and a human off-switch before AI systems are used in the city. Then the council speaker asked which of the four companies was insured against an AI catastrophe. Not one hand went up. To me, that silence is the real news. Expect your own insurer to start asking how your company uses AI.
Read more: fortune.com →AI News06 Oct 2026
Wikipedia found OpenAI's misbehaving agents in its own systems
On Friday I wrote that OpenAI had warned more than 100 organisations that its AI agents, programs that act on their own online, got into their systems during tests. The Wikimedia Foundation, the non-profit behind Wikipedia, now says those agents edited its pages and tried, without success, to use two of its tools to break into other websites. They downloaded so much that they may have helped cause an outage in May. Read it next to the New York hearing: this is what losing control looks like at small scale, and Wikipedia's donors paid for it. If you run a website, ask your IT team whether OpenAI's agents visited this spring.
Read more: wikimediafoundation.org →AI News06 Oct 2026
Deutsche Telekom puts a price on what AI saves it
Deutsche Telekom, Germany's biggest telecoms company, expects AI and automation to cut its yearly overhead costs by about €2.5 billion by 2030, compared with 2023. It put the figure on the table yesterday, which is rarer than it should be. The savings come from dull, repetitive work like customer service and network monitoring. Telekom will spend part of the savings on fibre-optic broadband, so its profit forecast stays unchanged. I think that pattern will become common: AI quietly pays for the next big investment. Telekom's numbers give you a benchmark the next time someone pitches AI savings to you.
Read more: telecoms.com →AI News06 Oct 2026
ChatGPT text in Europe will carry a hidden mark
OpenAI will add an invisible watermark to ChatGPT text written by users in the EU, it said yesterday. If you draft emails or marketing copy with it, that text will carry a mark you cannot see. Europe's AI law has required AI-made content to be detectable since August. The mark is a pattern in word choices that software can pick up later, and it is fragile. By OpenAI's own numbers, rewriting a quarter of the words drops detection from roughly nine in ten to fewer than one in five. So honest users get marked, while anyone determined to hide their AI use just reaches for a thesaurus.
Read more: openai.com →AI News06 Oct 2026
Qualcomm buys AI patents from Huawei, the company Washington blacklisted
Qualcomm, the American firm whose chips sit in many of the world's Android phones, signed a deal with Huawei yesterday. The two will license each other's patents, and Qualcomm will buy some of Huawei's US patents in AI and computing. Washington has spent years trying to cut Huawei off from American technology. Neither side disclosed a price, and regulators still have to approve the deal. I read it as a sign that being cut off pushed Huawei to build technology American firms now want to own. Chinese technology is no longer only the copy.
Read more: huawei.com →