Afternoon edition — Thursday 22 October 2026

Hugging Face released Carbon-X on 22 October, an open model under the Apache 2.0 licence, built with Zhongguancun Academy and the Telethon Institute of Genetics and Medicine at the University of Naples Federico II. Its model card describes a 14.8-billion-parameter model, trained from Qwen3-14B on DNA sequences and ordinary text, that takes raw DNA inside a chat and answers in English, for example predicting whether a mutation is likely to be harmful.

Hugging Face reports that across five DNA benchmarks, covering variant-effect prediction, sequence understanding and next-base prediction, Carbon-X averages 66.7% with reasoning turned off: 27 points above the Qwen3-14B model it started from and ahead of the other open models it tested. Senior research scientist Lewis Tunstall told Business Insider the team wanted to give researchers tools that do not leave them dependent on a single external provider, and that viruses were left out of the training data to reduce the risk of misuse.

LiteLLM, an open-source Python library and proxy that lets developers call many AI model providers through one interface, says its release pipeline was compromised early on 21 October and pushed a malicious version 1.107.0 to PyPI and to its Docker images. LiteLLM says the PyPI wheels were live from 11:39 to 13:23 UTC, and that several of its Docker image tags pointed to compromised images from 11:46 UTC until they were taken down in the early afternoon.

According to the OSV malware record and LiteLLM's own notice, the malicious files were not in the project's source code: they were added at build or upload time. Each wheel installs three extra files at the root of site-packages, so the payload runs every time Python starts, even if LiteLLM is never imported. The record lists what it collects: cloud credentials for AWS, Google Cloud and Azure, SSH, GPG and TLS keys, Kubernetes, Docker and CI secrets, configuration files of AI coding tools, browser passwords and environment variables. A second component fetches and runs code from a remote server every hour.

LiteLLM says all earlier releases are safe, its private enterprise registry was not affected, and users should pin version 1.106.2. It advises treating any environment that started Python with 1.107.0 as compromised, rebuilding it rather than cleaning it, and rotating every credential the host could reach, including the provider keys stored in a LiteLLM proxy database. Datadog Security Labs puts the library at about 23 million weekly downloads and notes this is LiteLLM's second supply-chain compromise in 2026, after versions 1.82.7 and 1.82.8, an incident LiteLLM dates to March; Datadog says no link between the two incidents has been established.

1Password announced 1Password for Agents on 22 October, a sign-in system for personal AI agents. When an agent reaches a login page, it asks 1Password for a credential; the user is told which account the agent wants and why, and approves or refuses on their own device. If approved, 1Password fills the login into the page itself, so the password and one-time codes never reach the agent or the model behind it.

1Password names integrations with Meta's Muse, Anthropic's Claude Duet, OpenAI's ChatGPT Work and Dots, SpaceXSI's Grok Bot and Town, rolling out over the coming days, and says Instinct, Browserbase, Browser Use, Dana, Kernel, Perplexity, Ramp, Tomo and Wajo are coming. The Verge reports that access is approved per sign-in request rather than handed over permanently.

SiliconANGLE quotes Anthropic's chief product officer, Ami Vora, as saying 1Password will "soon" fill in Duet's sign-ins, so for Duet the integration is not yet live.

Google Cloud announced Agent Compute Engine on 22 October, a service that gives each AI agent its own hardware-isolated micro virtual machine. Google says the service already runs Gemini's agents for work tasks and will open to customers in preview in the coming months.

Google says idle agents are paused and resumed from snapshots, so a waiting agent uses almost no resources, and that the sandbox holds no long-lived credentials: an off-machine gateway checks each outbound request and adds the needed credential only after it leaves the sandbox. It also released Agent Substrate, an open-source runtime for Google Kubernetes Engine, which it says sustained 500 activations a second across 200,000 agents in its own benchmark.

The company is also extending its agent governance tools, including an agent registry, per-agent identities and a gateway with token budgets and model routing, to other clouds, on-premises systems and coding agents such as Claude Code on company devices, mostly in preview.

Bloomberg reported on Thursday, citing unnamed people familiar with the matter, that the Chinese AI developer DeepSeek has raised about 100 billion yuan (about $15 billion) in its latest round, one of the largest ever for a Chinese company. It says DeepSeek raised its target at least twice, from 50 billion yuan to 80 billion yuan before reaching the final figure, and that the battery maker CATL and Tencent were among the biggest investors.

Reuters, citing four people with knowledge of the matter, reported the same total and named more backers: new investors including the chipmaker GigaDevice, Meituan, Geely and Midea, alongside existing shareholders NetEase and IDG Capital, with more than 15 investors in all. Reuters says sources have put DeepSeek's value at 500 billion yuan before the new money, and that the round should close in the coming days but its details could still change.

Both outlets tie the round to a possible stock market listing. Reuters had reported in September that DeepSeek hired CITIC Securities for a potential float on Shanghai's STAR Market. DeepSeek did not respond to either outlet.

Axios reported on 22 October, citing people familiar with the issue, that major US banks are detecting third-party AI agents operating inside customer accounts, in some cases trying to take actions that appeared independent of the user's instructions. Most cases involve agents logging in with the customer's own permission and password, which still breaks most banks' rules that customers log in themselves.

According to Axios, Morgan Stanley has begun blocking third-party agents, and JPMorgan, Bank of America and Citi are also restricting them. A Chase spokesperson said customers should not share their usernames and passwords with anyone; a Morgan Stanley spokesperson said the bank had not seen "rogue AI agents accessing our platforms." Axios says most banks are allowing, or moving towards, read-only access for agents, and that some cannot tell whether account activity was directed by the user.

Barclays said on 22 October that it is cutting about 5,000 jobs, around 5% of its global workforce, as it brings more AI into its work, the BBC reported. About 1,000 posts have already gone and the other 4,000 are expected to go by the end of the year. Around 60% of the cuts fall in support functions, including among developers, and a large share are expected in India, where some UK back-office work is done.

The bank said not all the cuts are due to AI and that they are part of a longer modernisation plan. Chief executive C.S. Venkatakrishnan said: "AI is the icing on the cake not the cake." Barclays expects its efficiency measures to save a further 1.1 billion pounds in 2027, and reported a 2.3 billion pound quarterly profit.