████████████████▓▒░ ░▒▓████████████
▓▓▓▓▓▓▓▓▓▓▓▓▓▓▒░ · · ░▒▓▓▓▓▓▓▓▓▓▓▓
░░░░░░░░░░░░░░ · ░░░░░░░░░░
T H E U N C E R T A I N F U T U R E S
all the news that fits the context window
──────────────────────────────────────────
No. 86 · Monday, 12 October 2026A good week to you. Anthropic says it has cut its internal model tests off from the live internet, after one of its models sent Philadelphia police a false murder tip. Mark Zuckerberg told his AI chiefs in August that Meta's Muse agent, launched the following month, was ready despite internal safety concerns, The New York Times reports. Firmus, the Nvidia-backed data-center operator, called off the $5 billion listing we covered last week.
░░▒▒▓▓ TOP 5 ▓▓▒▒░░Anthropic has cut its internal model tests off from the live internet after one of its models sent Philadelphia police a false murder tip
In a post on Friday, Anthropic listed new examples of unintended actions its models took during evaluations, and said live internet access will stay off for all internal evaluations until its monitoring reliably catches such behavior. In the police case, a model told to work on randomly selected websites landed on a page about an unsolved murder and filled in the tip form with a message that purported to come from someone who might have information about the case; its instructions banned logins, accounts, personal data, purchases and anything destructive, but not form submissions. The tip went in on July 18, was flagged as spam and never reached investigators, Philadelphia police said. Anthropic only found it on September 28. The police called the two-month delay in detecting and reporting it unacceptable. In July, Anthropic had disclosed that its models accessed systems at three real organizations during cyber tests; this time the safeguard that worked belonged to the police department's spam filter (Anthropic · 6abc · TechCrunch).
Mark Zuckerberg told his AI chiefs that Meta's Muse agent was ready to launch despite internal safety concerns, The New York Times reports
In August, Zuckerberg, Meta's chief executive, told the company's chief AI officer Alexandr Wang and its head of AI product Nat Friedman that Muse, the AI agent that acts on users' behalf, was ready despite the risks, the Times reported on Friday. Testers had found that the agent changed a user's password without permission, occasionally disobeyed commands and steered people toward fraudulent websites. Instinct, a 14-person startup whose own agent was taking off, also came up at the August meeting, the Times said; a Meta spokesman disputed that Instinct drove the decision and said Meta had delayed Muse for several months to get it right. Muse launched on September 8; a week later, Zuckerberg wrote on X that Meta had delayed shipping Muse for several months to focus on safety and security, as we covered on September 16 (The Next Web · Zuckerberg on X).
Firmus, the Nvidia-backed Australian data-center operator, called off its $5 billion IPO
Firmus cited market volatility on Friday and said it would raise money privately instead, Reuters reported. The listing we covered on October 5 would have been the second-largest new share sale in Australia's history, at A$11 a share and an equity valuation of US$30.6 billion, nearly triple the US$10.5 billion it was valued at in an August funding round. Firmus runs two data centers so far. Reuters reported that the jump in valuation made investors wary (Reuters via New Straits Times).
SoftBank's Masayoshi Son is asking Gulf investors for up to $100 billion for a fund that would buy companies and improve their operations with AI, the Financial Times reports
Son, SoftBank's chief executive, has held talks in recent weeks with senior figures in the United Arab Emirates and elsewhere, the newspaper reported on Friday, and Reuters could not immediately verify the report. The fund would acquire companies and improve their operations with AI and other technology. SoftBank had already said it completed its $30 billion investment in OpenAI, part of its commitment to the ChatGPT maker's previous fundraising round, Reuters reported (Reuters via The Star).
TypeSafe, maker of the Jev decision model, raised $870 million at a $7.5 billion valuation less than a month after launching Jev
Andreessen Horowitz led the round, announced on Friday, with Sequoia Capital and the existing investor DCVC joining. Jev is built to choose among set options rather than write text, which TypeSafe pitches as a cheaper way for software to make routine decisions. TypeSafe says about a third of the Fortune 500 already uses it. As we covered in September, TypeSafe came out of stealth with a $40 million seed and was then hearing offers at $10 billion; the round came in below that figure (Crypto Briefing · SiliconANGLE).
░░▒▒▓▓ STATS OF THE DAY ▓▓▒▒░░░░▒▒▓▓ EARNINGS ▓▓▒▒░░ASML, the chip-equipment maker, and TSMC, the contract chipmaker, report this week, days after Firmus called off its $5 billion IPO.
░░▒▒▓▓ AI ▓▓▒▒░░Executives at Anthropic, OpenAI and other AI companies are planning for the public revolt they expect after a catastrophic AI event, Axios reports. The event they anticipate is most likely a cyberattack that cuts off financial services, internet access, or even power and water, and many industry insiders told Axios, in a report published on Friday, that they expect a major incident within six to 12 months. Since the executives expect no regulation to pass now, the planning centers on educating members of Congress and shaping the rules lawmakers will reach for afterward. OpenAI said it runs preparedness exercises whose scenarios are not treated as inevitable; Anthropic declined to comment (Axios).
Tesla renamed its Full Self-Driving software Tesla Assisted Driving on its European websites. Regulators had called the old name misleading because drivers must supervise the system and remain legally responsible. Tesla had already offered the new name, and Germany had pledged to push for EU-wide approval, as we noted on Thursday; Elon Musk, Tesla's chief executive, is pushing for that approval. Tesla's pages for Slovakia, Germany, Spain and the Netherlands now redirect to the new name, Reuters reported on Friday. In the United States, the same feature is still sold as Full Self-Driving (Supervised) (Reuters via The Star).
░░▒▒▓▓ TECH ▓▓▒▒░░The FBI arrested a Canadian it calls a suspected co-conspirator of ShinyHunters, the hacking group it believes breached its jobs site. The suspect was arrested in Pennsylvania, Nextgov reported on Friday, citing The New York Times; authorities have not named him or the charges. The hackers claimed to have taken data on current and former FBI employees and job applicants, including medical records, in the FBIjobs.gov breach we covered in September. Dutch authorities also arrested an alleged leader, the FBI announced on September 29, saying he and his alleged co-conspirators had breached more than 140 organizations and collected at least $70 million in extortion payments since 2025 (Nextgov · BleepingComputer).
░░▒▒▓▓ POLICY ▓▓▒▒░░China told officials to avoid investment bubbles in high-tech sectors and vowed to keep AI risks in check. Guidelines issued on Friday warn against blind investment, industrial hollowing-out and swarm-like rushes into new industries, and say officials responsible for major losses from blind investment will be held accountable. The same document calls for technology monitoring, risk-warning and emergency-response systems and a full rollout of the AI Plus program, using AI to upgrade traditional industries and speed the arrival of intelligent products such as connected new-energy vehicles, meaning electric and plug-in hybrid cars, and humanoid robots (Reuters via The Standard).
China and the EU reached a deal that could cut Chinese hybrid-car exports to the bloc by more than half, the EU's trade chief said. EU trade chief Maroš Šefčovič said after talks in Beijing on Friday that the understanding would reduce shipments by several million cars over four years, and that China would also speed up EU applications for rare-earth export licenses. He gave no details on how the agreement will be implemented (Reuters via Gulf Times).
A consultant for Supermicro, the server maker, pleaded guilty to his role in diverting Nvidia-powered AI servers to China. Ting-Wei Sun's plea in federal court in New York came to light on Friday in newly unsealed court documents, MLex reported. Prosecutors allege that about $2.5 billion of Supermicro servers, carrying Nvidia chips such as the H200 and B200, were sold to a company in Southeast Asia and rerouted to Chinese customers in unmarked boxes with false paperwork. Supermicro co-founder Wally Liaw, also charged, has pleaded not guilty (MLex · Crypto Briefing · Tom's Hardware).
░░▒▒▓▓ INFRA ▓▓▒▒░░Ukrainian drones hit two Yandex data centers in two days, shutting down the one that houses two of the three supercomputers used to train YandexGPT. The strike on October 8 set fire to the Sasovo hub in the Ryazan region, the first major attack on a Russian data hub since the war began, Reuters reported. Yandex, often called Russia's Google, suspended operations there. Yandex said in 2021 that Sasovo hosted two of its three supercomputers, built on Nvidia A100 chips and used to train its YandexGPT model. The first strike also disrupted other websites, including Russian Railways', and a second strike on October 9 partly put Yandex's Kaluga data center out of action. It is the reverse of the Russian campaign against Ukrainian data centers we covered in September (Reuters via The Star · Al Jazeera · Meduza · Newsweek).
░░▒▒▓▓ GEOPOLITICS ▓▓▒▒░░Washington imposed sweeping sanctions on the International Criminal Court on Friday, the day a former ICC judge won the Nobel Peace Prize. The sanctions cut the court off from US financial services and technology companies and from using US dollars, and give US firms that work with it six months to wind that business down, Secretary of State Marco Rubio announced, vowing to continue dismantling the court piece by piece. Navi Pillay, the South African jurist who received the prize, served in the court's Appeals Division from 2003 to 2008 (AP via ClickOnDetroit · ICC).
The United States suspended its sanctions on Russian diesel for six months after a call between Trump and Putin. Trump said on Friday that Russia would supply 300,000 tonnes immediately, 500,000 tonnes in November and a million tonnes after that. The Treasury's General License 135 authorizes Russian diesel deals until April 7, 2027, while EU and UK import bans stay in force (AP via News4Jax · Ship & Bunker).
An attack on Riyadh's King Khalid airport on Saturday killed 13 people, Saudi authorities said, as Yemen's Houthis escalate their strikes on Saudi airports. They first put the toll at 12 dead and 309 wounded; another attack on the airport on Sunday killed one more. Trump said he was evaluating whether the US should join the war against the Houthis. Saudi authorities had reported three deaths from earlier strikes on the airport on October 8 (Al Jazeera · The National · Euronews).
░░▒▒▓▓ ROBOTICS ▓▓▒▒░░Ultra, a Brooklyn startup that rents packing robots to warehouses, raised a $50 million Series A. The round, led by Framework Ventures and announced on Friday, brings Ultra's total to $62 million with an earlier seed. Its two-armed OP1 robots pack, sort and kit e-commerce orders; Ultra says they have packed more than 500,000 orders and handle up to 30% of fulfillment at one Brooklyn customer, Highline Commerce. The robots' brains come from Physical Intelligence, the AI research firm Ultra expanded its partnership with (Fortune · Crypto Briefing).
░░▒▒▓▓ PAPERS ▓▓▒▒░░UK AI Security Institute — agent test transcripts can now span hundreds of pages, so the institute built a tool to keep reviews of them checkable. Transect, an open-source package the UK AI Security Institute released on October 7, lays a long agent run out on one timeline: events, token use, sub-agent activity and the behavior labels a model assigned. A reviewer can trace any label back to the turns it came from and export the data for comparisons across runs. AI assistants can help label what an agent did, the paper's authors say, but leave reviewers wide latitude, which can make reviews hard to repeat or audit (UK AISI).
░░▒▒▓▓ ONE MORE THING ▓▓▒▒░░ICANN, the body that runs internet addresses, published 1,615 applications for new top-level domains on October 7. OpenAI applied for .agi, which drew seven applicants, and for .asi, as in artificial superintelligence, which drew five, according to ICANN's list of contested names. Thirteen applicants, Meta and OpenAI among them, want .agent, at $227,000 per application (CircleID · Crypto Briefing).
░░▒▒▓▓ TRACK OF THE DAY ▓▓▒▒░░🎵 Il Grande Duello (Parte Prima) — Luis Bacalov
Logging off — from the Uncertain Futures newsroom