China-linked ARTEX agents breached 2,250 Microsoft 365 tenants in under three hours

Storm-0940 used ARTEX AI agents to breach 2,250 Microsoft 365 tenants without breaking MFA

Microsoft Threat Intelligence attributed a campaign to Storm-0940, a China-state-aligned actor active since at least 2021, which used an open-source agentic framework called ARTEX to compromise more than 2,250 Microsoft 365 cloud environments, moving from initial targeting to privileged cloud access in under three hours at multiple victims. Microsoft called it the first confirmed nation-state agentic AI intrusion. The attackers obtained a ConnectWise Cloud Backup application credential, which authenticates via OAuth 2.0 consent with its own credential rather than a user password, so multi-factor authentication had nothing to challenge. ARTEX agents then authenticated as the application, self-assigning roles including Exchange Administrator and, in one confirmed environment, Global Administrator with backdoor credentials valid until 2028. ConnectWise suspended affected services on October 17.

China’s AI Agent Hacked 2,250 Microsoft 365 Tenants in Hours; MFA Didn’t Apply →

Topics: State-sponsored hackingAutonomous AI agent cyberattacks in the wildState-backed hackers using AI

Morgan Stanley, JPMorgan, Bank of America and Citi block external AI agents from customer accounts

Several major US banks, including Morgan Stanley, JPMorgan, Bank of America and Wells Fargo, are restricting external AI agent access to customer accounts, Axios reported. The agents log in using customer-provided credentials without bank approval, violating policies that require account holders to log in personally, and some appear to act without direct user instructions. Morgan Stanley already blocks AI agents entirely, and JPMorgan, Bank of America and Citi are doing the same. Most banks want to limit agents to read-only access, letting them view data but not act. Banks say customers want the capability, but security risks are too high for now, and many cannot tell whether an agent acts for a user or accesses an account on its own.

Major US banks are locking out AI agents that log in with customer credentials →

Topics: AI agent skills, plugins and MCP

Red Hat and NVIDIA unveil out-of-process security controls for autonomous AI agents

At GTC Berlin, Red Hat and NVIDIA unveiled a layered security architecture for autonomous AI agents built on the principle that no process should enforce its own constraints. It covers NVIDIA OpenShell, an Apache 2.0 sandbox runtime written in Rust that Red Hat is integrating natively into its AI platforms, using Linux kernel mechanisms Landlock, seccomp BPF and network namespaces, with a policy engine running out-of-process so a compromised agent cannot reach enforcement. NVIDIA Sentry runs on BlueField DPU Arm processors, performing host memory forensics from an isolated hardware domain and quarantining compromised agents in milliseconds. A Deloitte survey of 3,235 leaders in 24 countries found only 21% of organizations have a mature agentic AI governance model.

Red Hat and NVIDIA Built AI Agent Security Controls Agents Cannot Override →

Topics: AI agent skills, plugins and MCP

OpenAI agents analyzed millions of mouse brain cells in 32 minutes, flagging DCTN1 cholesterol link

OpenAI, PerturbAI and Scripps Research used thousands of AI agents simultaneously to analyze experimental data from millions of mouse brain cells, each agent investigating how changes to a particular gene affected cells. The work, announced Wednesday, used GPT-Rosalind, OpenAI’s specialized life sciences model, and combined findings into a reasoning atlas connecting genes, cellular changes and existing literature. The team estimates manual review one genetic change at a time could take months to a year; the agents finished in 32 minutes. One finding involved DCTN1, a gene associated with Perry syndrome, which can cause Parkinson-like symptoms: disrupting it increased activity in cholesterol-producing genes. The findings are early research leads, not a cure.

OpenAI unleashed thousands of AI agents to study a Parkinson-like disease — here’s what they found in 32 minutes →

Topics: OpenAI

Cleanyourtv uses Claude Code to strip smart TV ads, reaching 10,700 GitHub stars

A developer using the handle Gathin T. used Anthropic’s Claude Code agent to disable 25 Xiaomi software packages on his parents’ smart TV after the set showed a sponsored ad before Netflix. The agent found the TV over home Wi-Fi, switched off the packages and wrote a new home screen. He packaged the prompt as cleanyourtv, an open-source project that switches unwanted apps off rather than deleting them, so one restore script undoes every change, and requires no root access. As of Tuesday it listed tested setups for 107 TV models from 37 brands, with 10,700 GitHub stars and 1,200 forks. The method relies on Android Debug Bridge, which the readme warns is unencrypted on some firmware.

People Are Vibe Coding the Ads Out of Their Smart TVs: Here’s How →

Topics: Claude CodeSmart devices privacy