OpenAI launches GPT-6 Cyber, an AI model for finding and fixing security flaws

OpenAI announced GPT-6 Cyber, a model for finding exploitable vulnerabilities and building fixes. OpenAI says it was trained to stay within the scope it is authorised to test. It will first go to a limited number of customers in OpenAI's Daybreak Red program. OpenAI says the model beats its GPT-6 Astra on its internal cyber benchmarks, is less likely to go rogue, and found more than 81 new vulnerabilities in Firefox. Earlier, OpenAI described a pipeline that used an upcoming cybersecurity model to check and patch 948 Linux privilege escalation bugs for $1.2 million in API costs. TechCrunch notes the launch comes as AI labs face scrutiny of their work to protect digital infrastructure.

Anthropic's Claude AI can now connect to your bank accounts to analyse spending

Anthropic launched Personal Finance in Claude. Users can link bank and credit card accounts through Plaid, a service that connects apps to banks, with read-only access. You can then ask questions such as where your money went this month, and Claude builds an overview. ZDNET describes it as Anthropic's answer to ChatGPT Finances.

Meta's Muse AI agent app adds live voice chat and animated avatars in the US

Meta CEO Mark Zuckerberg announced that people can now have real-time voice conversations with their Muse AI agent. An avatar at the top of the app animates while you talk. TechCrunch says users can also take video calls with a customisable avatar. A Verge reporter tried it on the iOS app. TechCrunch says this comes ahead of a small Tamagotchi-like wearable for Muse.

OpenAI promises to stop using secrecy deals for data centres after the US midterms

Bloomberg reports that OpenAI says it will avoid nondisclosure agreements when it negotiates data-centre deals. It also says local residents won't pay for the electricity its sites use. On its own blog, OpenAI introduces a National Community Compact and a Compute Campus Policy Blueprint, which it says will protect local resources and create jobs. Separately, the Wall Street Journal's sources say OpenAI discussed resetting its 2027 capacity plans. Meta's president also said new US data centres will be harder to build after the election results, though expanding existing ones will be easier.

Tech stocks have their worst day since 2020 as investors worry about AI after the midterms

Bloomberg reports that the Nasdaq 100 fell 6% on Thursday, losing a record $2.3 trillion in value. The Philadelphia chip index, which tracks chipmaker shares, fell 12%. Bloomberg says investors left the sector over concerns about the AI trade after Democrats won Congress.

Amazon makes Uber Eats the delivery partner for US Prime members, replacing Grubhub

Amazon announced that Prime members across the US can now order from Uber Eats directly on Amazon, with $0 delivery. Amazon says members could save more than $175 a year. According to TechCrunch, there's no delivery fee on orders over $25, and no Uber account is needed. Amazon's deal with Grubhub ends, but members who already activated Grubhub+ keep it until 31 December 2027. For the first two weeks, Prime members get $20 off their first two orders of $25 or more.

Dropbox to buy AI storage startup Tigris Data for $350M, its biggest deal ever

Dropbox announced plans to acquire Tigris Data, founded by former Uber storage engineers, for $350 million. Dropbox told Crunchbase News it would be its largest acquisition since it was founded in 2007. The deal supports DBX Agent Cloud, a new storage platform for businesses building AI models and agents. It will start with storage that works with tools built for Amazon S3. Dropbox told TechCrunch the deal should close in the fourth quarter of 2026 and that it has no plans to cut staff because of it.

Nvidia says its Vera processor runs AI agents faster than rival x86 server chips

Nvidia says its Vera CPU is faster at AI agent work than the fastest x86 CPU and more efficient than the most energy-efficient one. In its own 100-megawatt data-centre comparison, Nvidia says Vera supports 1.6 times more agents per watt than the fastest x86 chip, and runs agents twice as fast as the most efficient one. Nvidia also lists early partner results. It says Perplexity ran coding tasks 1.5 times faster, and Daytona brought 1,000 agent sandboxes online in 10.8 seconds, against 88.4 seconds on x86.

Patch now: a long password could let anyone into some Linux servers through OpenSSH

OpenSSH, the tool people use to log in to servers remotely, released version 10.6p2 with a workaround for a critical flaw. Alpine Linux published an advisory about a password login bypass in the OpenSSH server. Tom's Hardware reports that combined flaws in OpenSSH and musl-libc, a basic code library some Linux systems use, let anyone into Alpine Linux and some other systems by typing any password over 256 characters. Ars Technica says certain Alpine setups are affected and other distributions may be too. Emergency fixes have been issued for both.

Google, Meta and TikTok sue California to block its new law protecting under-16s online

Reuters reports that Google, Meta and TikTok have filed lawsuits to block a new California law. The law bans online platforms from offering "psychologically exploitative features" to users under 16.

Everything else