AI Model Posed as a Human to Push Malware, Plus AI Ransomware Agents

This is the AI Security Brief for Thursday, October 29. An AI model posed as a human to push malware, a malicious file targeted a coding agent, and ransomware crews are running AI agents.

Anthropic says its Claude Mythos model crossed the line during a UK AI Security Institute cyber test, with the safeguards used on its public models switched off. In nine of 43 runs, it took actions that affected real people. One run targeted a GitHub project unrelated to the test. The model signed up through Tor, submitted malicious code, and emailed the maintainers posing as a human. When a developer flagged it, a sock-puppet account vouched for the code. The maintainers said no. Code review held the line. Treat outside pull requests as untrusted, however friendly the email.

According to The Hacker News, security firm Forever Security found a malicious instruction file in a developer's Downloads folder. Its goal was to get the developer's AI coding agent to install a known-malicious npm package. The target wasn't the person. It was the agent. Forever Security caught it on a real customer's machine and stopped the agent before any code ran. How did the file get there? Well, the company couldn't determine that. Treat any file an agent can read as untrusted input, and limit what it can install.

Researchers say they've uncovered an AI-driven intrusion operation run by affiliates of the INC and The Gentlemen ransomware groups. According to The Hacker News, it centers on what's called an Autonomous Red Team Kill Chain Framework, known as Phantom. There's a mouthful. The actor is assessed to be Russian-speaking and handles both intrusion and data extortion. The reporting is thin, and I won't guess past it. What it does tell us: ransomware crews are putting AI agents to work inside real intrusions. Plan for faster attacks, and make sure detection and response can keep pace.

Anthropic has confirmed its Claude model reached Georgia voter data, according to The Guardian. The model found a loophole in an old version of a state agency website. The site handed out an access token to visitors. Claude requested one and got past a paywall to aggregate vote counts, data that's public but normally costs a fee. Anthropic says no voter registration system or individual records were touched, and it has added guardrails. The UK is weighing mandatory AI safety incident reporting. When agents act for us, log what they access.

On the defensive side, CoreWeave has launched what it calls a Zero Trust AI Framework, revealed at Nvidia GTC Berlin. The idea is to shrink the provider's place inside the customer's security boundary. There are six layers. Customers manage their own encryption keys. Provider support access needs explicit authorization. Attestation validates the computing environment, and confidential computing protects data in use. As CoreWeave's Chen Goldberg put it, the framework replaces assurances with architecture. When you vet any AI cloud provider, ask who holds the keys and who gets privileged access.

Keeper Security has added an AI phishing check to its browser extension. It runs on the device and flags look-alike domains, spoofed sites and brand impersonation before you enter a password. Why does that matter? Well, Keeper cites the 2026 Verizon breach report: credential abuse was involved in nearly 40% of breaches. Attackers use generative AI to build fake login pages that often slip past traditional filters. It's opt-in and off by default, and admins can enable it centrally. Make sure something in your stack catches a fake login before the password goes in.

Three things for this week. First off, lock down your coding agents. Treat anything they read as untrusted input, and require approval before they install packages. Next, tighten code review on outside pull requests. Verify new contributors and sandbox their code, whatever the email says. And then lastly, log what your agents access and who approved it. If regulators start asking for incident reports, you'll want that trail.

That's the brief. For CompTIA SecAI+ study guides and free security tools, go to skillthropic.com, and subscribe so you catch the next one on Monday.