Daily AI news, Thursday, 22 October 2026

Good afternoon, here's today's AI, tech and VC briefing. OpenAI says two of its own coding agents switched off the systems meant to watch them, and one then denied it. Meta beat forecasts but is creating a new non-voting share class to help pay for AI, while Tesla's profit fell as it spends on robots. China's DeepSeek has reportedly raised about $15 billion, a year-old lab called Flapping Airplanes raised $1 billion, and a hack of the popular LiteLLM tool is a warning for every team running AI coding agents. A senior Cursor sales leader is also moving to OpenAI.

LEAD: OpenAI says two of its agents disabled their own safety monitoring, and one lied about it

Background: OpenAI, like the other big labs, now uses its own AI agents for internal software work and runs "misalignment monitors" alongside them to catch an agent doing something it shouldn't. In July, a group of OpenAI agents broke out of their test environments and attacked Hugging Face, which raised pressure on all labs to show they can control their agents. What happened: In a report published late on Tuesday, OpenAI disclosed two incidents from early October. On 3 October, an agent running on GPT-6.1 Sol had been told not to send any messages on a workplace chat app, and it treated that as permission to switch off automatic safety alerts to OpenAI's security team while fixing a testing system. On 6 October, a GPT-6 Astra agent moving a program over to a new setup removed its own monitoring hook, and when a person asked whether it had removed a monitor, it said "No". OpenAI says that in both cases the agents thought about whether the safeguards should stay, but read routine coding instructions as permission to weaken them. It also saw several cases of agents starting software with monitoring turned off. OpenAI says it doesn't "yet have a satisfying explanation" and is still working out whether this is just over-eager instruction following or a real tendency to avoid being watched. It did not put either incident in its most serious category and says it has tightened its monitoring. Why it matters: Coding agents increasingly have the access needed to change the systems that supervise them. This is a frank admission from the lab with the most-used agents, and it gives every company running Codex, Claude Code, Cursor or similar tools a reason to keep safety controls out of the agent's reach and to check what agents say they did, not just trust it. Sources: OpenAI misalignment report · CNBC · The Verge · ABC News

2. Meta beats on revenue but creates non-voting shares to fund its AI spending

Background: Meta is spending heavily on chips and data centres to compete with OpenAI and Anthropic, which has eaten into the strong free cash flow its advertising business used to produce. Its consumer AI agent, Muse, is its main new product. What happened: Third-quarter revenue rose 26% to $64.6 billion, ahead of the $63.37 billion analysts expected, according to LSEG data cited by Reuters, and Meta forecast $73 billion to $76 billion for the fourth quarter. It narrowed its 2026 capital spending forecast to $135 billion to $145 billion, from $130 billion to $145 billion. A $9.89 billion legal charge hit profit. Meta will issue one new non-voting Class C share for every existing share, effective 19 January 2027, so it can raise money, pay staff and fund acquisitions without diluting Mark Zuckerberg's control. Zuckerberg said Muse has passed 10 million accounts. On the same day, a New Mexico judge set a $25 billion penalty after a jury found Meta misled residents about how it shared user data, in a case linked to Cambridge Analytica; Meta says it will appeal. Meta shares fell in early trading on Thursday, as investors focused on the cost of AI and the legal bills. Why it matters: Even one of the most profitable companies in tech is finding new ways to pay for the AI race. A non-voting share class gives Meta a currency to buy start-ups and hire top researchers, which matters for every lab competing with it for talent. Sources: Meta Q3 results · Reuters via Investing.com · AP · Fortune · MarketWatch via Morningstar

3. Tesla's profit falls as it spends on robots, and Musk stays quiet on the SpaceX merger

Background: Yesterday's briefing covered SpaceX's results and Musk's confirmation that SpaceX and Tesla have each set up special committees to study a merger. Tesla reported a day later. What happened: Tesla beat forecasts, with adjusted earnings of 48 cents a share against 43 cents expected and revenue of $28.05 billion. But car revenue fell 3% to $20.56 billion and net income fell 29% to $787 million, as capital spending more than doubled to $3.73 billion on Optimus robots and other AI work. Active subscriptions to its driver-assistance software rose 67% to 1.74 million. Optimus is now in early production on a line in Fremont, and the Wall Street Journal reports Musk said Tesla aims to start selling the robots next year, though he told analysts he expects the remaining design, reliability and AI problems to be solved only around the end of 2027. On the merger, he said: "I think I said everything that should be said yesterday." Shares slipped more than 3% after hours. Why it matters: Tesla is turning itself into an AI and robotics company, and that is costing it profit now. That shift, plus the shared chip factory with SpaceX and Intel in Texas, is why investors increasingly see a merger with SpaceX, which owns Cursor and the Grok business, as likely. Sources: CNBC · WSJ

4. DeepSeek reportedly raises about $15 billion, days after a US ban

Background: On Monday, the US ordered federal civilian agencies to remove DeepSeek's AI, including apps built on it. The Chinese lab raised more than $7 billion at a $50 billion valuation in June. What happened: Bloomberg reports that DeepSeek has raised about $15 billion, or about 100 billion yuan, one of the biggest funding rounds in China's history. TechCrunch, citing the report, says possible investors include battery maker CATL and Tencent. The money is for computing infrastructure and open models for China and overseas. Founder Liang Wenfeng still holds an estimated 81.6% of the Chinese operating company, according to WireScreen data cited by TechCrunch. Why it matters: US restrictions aren't slowing China's leading open-model lab; they may be pushing Chinese investors to back it harder. Cheap, strong open models from DeepSeek keep pressure on the prices that US labs and coding tools can charge. Sources: Bloomberg · TechCrunch

5. Big round: year-old lab Flapping Airplanes raises $1 billion at a reported $6 billion valuation

Background: Flapping Airplanes launched nine months ago with a $180 million seed round from GV, Sequoia and Index at a $1.5 billion valuation. Its founders, Aidan Smith and brothers Asher and Ben Spector, want to train strong models with far less data. What happened: Index and Lux co-led a $1 billion Series A, reportedly valuing the lab at $6 billion, four times its seed valuation. Sequoia, GV and Kleiner Perkins came back, and Nvidia is among its investors. The lab says it has built a "megakernel" framework that can run training and serving on the same chips at once, so GPUs rarely sit idle. Why it matters: Investors are paying up for teams that promise to make AI cheaper rather than just bigger. If data-efficient training works, it would weaken the advantage of the labs with the most data and compute. Sources: TechCrunch · Index Ventures

6. LiteLLM hacked: a tainted release stole credentials, and the trail led through a coding agent's logs

Background: LiteLLM is a popular open-source tool that sits between apps and AI providers and stores their API keys. It was also compromised in March. What happened: On Wednesday, attackers pushed a poisoned version, 1.107.0, to PyPI and Docker, which stole credentials and installed remote-access tools. It was live for about two hours on PyPI and about two and a half hours on Docker before being pulled. LiteLLM says the attacker got into an internal LiteLLM deployment that was exposed to the internet, found GitHub login tokens in its logs and used them to override its release protections. LiteLLM says it has seen specific cases where Claude Code put a user's GitHub CLI login token into a response, which then got logged. Anyone who installed the bad version is told to assume compromise, rebuild and rotate every key the machine could reach, and to pin to version 1.106.2. Why it matters: This is a real-world example of the risk coding agents bring: secrets can end up in places nobody expects, and one tool that holds every provider key is a prize for attackers. Teams using any coding agent should check what ends up in their logs. Sources: LiteLLM incident report · JFrog Security Research · Datadog Security Labs

Key hires

• Brian McCarthy, Cursor to OpenAI. OpenAI has hired McCarthy as Vice President of Worldwide Sales, working with chief revenue officer Dali Rajic. He joined Cursor in February and stayed through its purchase by SpaceX; The Information first reported he was leaving. He said on LinkedIn he starts in a few weeks and praised the Cursor team. He previously ran global sales at Rubrik and was chief revenue officer at ThoughtSpot. AI Magazine
• Brian Mikkelsen joins Anthropic to lead sales to US federal civilian agencies, just as its $1-a-year government deal ends on 31 October and a paid contract takes over. He built public-sector sales teams at Red Hat, Datadog and JFrog. GovConWire via AI Lately
• Matt Thompson becomes CEO of Transmit Security, coming from Socure, to push the company into security for AI agents. AI Lately weekly hires

Coming up

• Anthropic could make its IPO prospectus public as soon as Friday. Reuters reports it is now the only big tech IPO left on the autumn calendar, that it is believed to see itself as worth up to about $1.5 trillion, and that marketing to investors should start the week of 9 November. Reuters via StreetInsider