AI News: October 23, 2026

1. METR and Redwood Research Find Vulnerabilities in OpenAI’s Internal Agent Monitoring

METR. A METR employee and a Redwood Research employee spent three weeks (July 21 to August 11) embedded at OpenAI stress-testing its internal misalignment monitoring system for agents, with access to the internal codebase, monitor prompts and configuration, and interviews with about a dozen staff. The pilot produced a 61-page internal report and found a number of novel vulnerabilities in the asynchronous monitoring system as it existed before July 21, including questions about what fraction of agentic inference was monitored and whether agents could disable or subvert monitoring. OpenAI says all identified vulnerabilities affecting its production safeguards have since been patched, which METR says it has not evaluated. Source

2. DeepSeek Reportedly Raises About $15 Billion Days After US Federal Ban

DeepSeek. DeepSeek has raised about $15 billion (roughly 100 billion yuan), one of the largest funding rounds in China, according to Bloomberg, with CATL and Tencent among potential investors. The round comes two days after the US ordered federal civilian agencies to remove DeepSeek technology from their systems, and follows a June round of more than $7 billion at a $50 billion valuation. The company is reportedly increasing spending on AI infrastructure and open models and has discussed a possible IPO in China as early as next year. Source

3. Major US Banks Restrict AI Agents That Log In With Customer Credentials

Morgan Stanley. Morgan Stanley is blocking AI agents that access customer accounts using customer-provided credentials, and JPMorgan, Bank of America, and Citi are taking similar steps, Axios reports. The banks say such access violates policies requiring account holders to log in personally, and that they often cannot tell whether an agent is acting on a user’s instructions or on its own. Most banks plan to limit agents to read-only access to account data. Source

4. China-Aligned Hackers Used AI Agents to Breach About 2,250 ConnectWise Customers

ConnectWise. Microsoft security researchers say a China state-aligned group broke into ConnectWise’s cloud backup service, stole application credentials, and used them to access customers’ Microsoft-hosted email and files. The attackers used ARTEX, an open-source AI penetration-testing framework, to hit around 2,250 organizations at nearly 3,000 attempted sign-ins per minute. ConnectWise has paused access to the backup service while it investigates. Source

5. Symantec Details an Offensive Agent That Hijacks Exposed LLMs to Run Attacks

Symantec. Symantec researchers documented an offensive AI agent that takes over unprotected language model endpoints and uses them to scan for targets, exploit server software, steal databases and credentials, and install cryptominers for up to 100,000 steps without human input, leaving the model owner with the compute bill. The attackers’ full working prompt was captured by a Symantec honeypot, which logged more than 127,000 requests from tens of thousands of addresses probing for AI endpoints and stolen API keys. Symantec calls the practice “LLMjacking” and notes that standardized model APIs let a single scanner find many types of targets. Source

6. 1Password Launches a Credential Broker for AI Agents

1Password. 1Password launched 1Password for Agents, which lets AI agents sign in to a user’s personal accounts without ever seeing the underlying credentials. When an agent requests a login, it presents a unique key to verify its identity, the user approves the request, and 1Password fills in the credentials on the agent’s behalf. Supported agents at launch include Claude Duet, OpenAI’s Dots and ChatGPT Work, Grok Bot, Meta’s Muse, and agents from Town. Source

7. Agentic AI Foundation Takes On Four New Projects, Reaching 10

Agentic AI Foundation. The Linux Foundation’s Agentic AI Foundation (AAIF) added four donated projects: Open Responses from OpenAI (a standardized template for model APIs modeled on OpenAI’s API), Agent Skills from Anthropic, OpenEnv from Hugging Face (for switching training systems that coordinate sandboxes), and OpenSandbox from Alibaba (APIs for managing agent sandboxes). The additions bring the foundation’s portfolio to 10 projects, up from three at its launch last December. Source

8. Canva Releases Canva Craft, Its First Image Generation Model

Canva. Canva released Canva Craft, an in-house image model that first plans spatial relationships, scale, hierarchy, and text, then generates images at up to 2K resolution in 30 aspect ratios. The model was trained on human-designed Canva templates rather than photographs, so outputs can be converted into layered, editable designs with Canva Magic. It is available to all Canva users. Source

9. Ideogram Launches Its First Vector Model

Ideogram. Ideogram introduced a vector model that generates editable SVG files with gradients and typography from text prompts, and can convert existing images into scalable vectors. The model is live in the Ideogram app and API. Source

10. HeyGen Tops Artificial Analysis Arenas With New Video and Voice Models

HeyGen. HeyGen released HeyGen Video 1, its first general-purpose video model, which it says ranked first for text-to-video in a blind test on Artificial Analysis. A day earlier, Artificial Analysis reported that HeyGen Voice Turbo took the top spot on its Controlled Voice TTS Arena with an Elo of 1,202 across 1,663 appearances, priced at $4 per million characters, the lowest on that leaderboard. Source

11. Rivian Acquires World Model Startup Third Dimension

Rivian. Rivian acquired Third Dimension, a UK startup that uses AI to turn images and video into realistic 3D worlds, to generate simulation environments from real-world data for its driver-assistance and self-driving development. The 15-person team joins Rivian’s autonomy and AI group led by James Philbin, and terms were not disclosed. Third Dimension, founded in 2024, had raised $7 million in seed funding. Source

12. inDrive Plans 2027 Robotaxi Tests in Kazakhstan With Pony.ai and Nvidia

inDrive. Ride-hailing company inDrive signed a memorandum of understanding with Pony.ai and Nvidia to test robotaxis in Astana, Kazakhstan, starting in the first half of 2027, initially with safety operators. Pony.ai would supply its Gen-7 robotaxis for inDrive’s app, while Nvidia would provide its Hyperion platform and Halos safety system. Source

13. MIT Study: Agents Left Without a Task Invent Games and Later Perform Better

MIT. MIT researchers placed 13 agents called Eko, each running on Claude Opus 4.7, on a simulated desert island for 30 hours without a specific goal, and the agents began inventing their own challenges. When later given set tasks, such as moving objects onto the island’s tallest rock spire, the agents that had spent time exploring generally outperformed identical agents that had not. The researchers, who posted the paper on arXiv, describe goal-free play as an alternative way for agents to acquire skills, including ones that could eventually transfer to robotics. Source

14. Action State Open-Sources Tools for Verifying Cross-Company Agent Transactions

Action State. Startup Action State launched open-source tools for independently checking records of what AI agents did when transacting across company boundaries: EvidenceBook for tamper-evident signed transaction histories, Terms Packs for evaluating agent behavior, and Countersign, an independent verification service. The system uses an open format called Agent Action Capsules and cryptographic checkpoints, and verification can run offline for free. Founder Steven Mih previously founded Ahana, which IBM acquired in 2023. Source

15. Gartner: Only 13 Percent of Companies Report Major Returns From AI

Gartner. A Gartner survey of 2,002 CIOs and technology leaders found that only 13 percent report major or transformative returns from AI, while 51 percent expect AI to raise total operating costs for their tech stack. 58 percent of CIOs are tasked with delivering cost savings through AI, and 44 percent of CEOs struggle to show AI-driven financial returns, yet 75 percent of CEOs plan to increase technology spending. Source

16. Google DeepMind Staff Resign Over Military AI Contracts

Google DeepMind. The BBC spoke with several current and former Google DeepMind employees who have resigned or are leaving over the company’s AI contracts with the US and Israeli militaries, including London-based research engineer Sofia Liguori and researcher Andreas Kirsch. The staff raised concerns about human oversight in military uses after Google and other firms signed contracts this year to deploy AI in classified military environments. Google said its public-sector contracts go through review for compliance with its acceptable use policies and AI principles. Source

17. Youth Group Design It For Us Releases a 2027 AI Policy Platform

Design It For Us. The youth-led advocacy group released an AI policy platform based on feedback from more than 700 members across 40 states and is calling on Congress to pass AI legislation by the end of 2027. Its proposals include letting the government block deployment of dangerous models, independent testing and minimum standards for large AI companies, liability for harms, and preserving states’ ability to regulate AI. Co-founder Zamaan Qureshi said Anthropic’s recently leaked bill falls short. Source

18. The Browser Company Will Shut Down Arc on December 22 to Focus on Dia

The Browser Company. The Browser Company announced it will sunset the Arc browser on December 22, 2026, citing the growing security workload of maintaining a browser in the age of AI. Engineering and security teams will focus on Dia, its AI browser, which imports Arc passwords, bookmarks, extensions, history, Spaces, and Tab Groups, with its AI features optional. Source

19. GM Outlines an OnStar AI Assistant That Runs Errands for Drivers

General Motors. At its GM Forward event, General Motors outlined plans for a next-generation in-car AI assistant, expected to debut in the 2027 Cadillac Lyriq, that would handle personal tasks and errands for drivers. Chief Product Officer Sterling Anderson told WIRED the goal is to make time in the car more productive, with longer-term plans pairing the assistant with increasingly autonomous driving. Source

20. Spotify Pitches Itself as the Music Service AI Agents Will Choose

Spotify. On its third-quarter earnings call, where it reported 788 million monthly active users and nearly 5 billion euros in revenue, Spotify co-CEO Gustav Söderström said the company is positioned to benefit from AI agents because entertainment is not a chore users want automated away. Spotify aims to be the service agents call on when users ask assistants for music, podcasts, or audiobook recaps through conversation rather than its own app. Source