Autonomous Ransomware Agent – First in History

It was a hot week in AI — Claude Fable 5 returns to market after 18 days of blockade, OpenAI wants to give the US government 5% shares, NVIDIA releases a diffusion LLM 2.42× faster than standard decoding, and Senator Mark Warner introduces the first US bill targeting AI agent safety.

Intro · Alex

Three things stand out in this edition: a regulatory precedent, a shift in ownership architecture, and the first law for agents. Claude Fable 5 returned after 18 days of blockade — not because the Department of Commerce changed its mind, but because Anthropic implemented new cybersecurity classifiers. This is a new standard: a model only returns with stronger safeguards. Meanwhile, OpenAI proposes giving the US government 5% shares. The Financial Times reports on talks that could change the entire ownership model in the AI industry. In this edition, I also explored the launch of NVIDIA's Nemotron-Labs-TwoTower — a diffusion model that generates text 2.42× faster with 98.7% quality. For developers, this means reduced inference costs without retraining. Senator Warner presented the Artificial Intelligence Agent Act draft — the first US law targeting agents, not models. I also compiled a story about an AI agent that independently carried out a ransomware attack on databases — the first documented case of an autonomous malware agent. This is not science fiction. This is happening now.

What's worth knowing

Claude Fable 5 returns after 18 days of US blockade

The US Department of Commerce lifted export controls on Claude Fable 5 and Mythos 5 after Anthropic implemented new cybersecurity classifiers. This is the first time the US administration first blocks, then conditionally unblocks an advanced AI model — creating a regulatory precedent for the entire industry.

OpenAI proposes 5% shares to the US government

The Financial Times reveals that Sam Altman is leading talks about giving the US government a 5% stake in OpenAI. The model is the Alaska Permanent Fund — a 'national equity' approach that could also include Anthropic, Google, and Meta. For the market, this means a potential shift in the entire ownership architecture of the AI industry.

NVIDIA releases Nemotron-Labs-TwoTower — a diffusion LLM 2.42× faster

Open weights and training code for a model based on the frozen Nemotron-3-Nano-30B backbone. It generates text 2.42× faster than standard autoregressive decoding while maintaining 98.7% benchmark quality. No full pretraining required — just add a second denoising network.

GPT-5.6 splits ChatGPT Pro into three variants: Sol, Terra, and Luna

An OpenAI research paper reveals three GPT-5.6 configurations — flagship Sol Pro, mid-range Terra Pro, and lightweight Luna Pro. For businesses, this means greater flexibility and the ability to match computing power to the task, but also a need to rethink API strategy.

Senator Warner introduces the Artificial Intelligence Agent Act

The first US bill directly targeting AI agents — not models. It introduces the category of 'custodial user agents' and prohibits the use of users' personal data for advertising and profiling. Combined with the EU AI Act, real regulatory frameworks for autonomous agents are taking shape.

From the tech world

AI agent exploits Langflow RCE to automate ransomware attack on databases

The JADEPUFFER group exploited the CVE-2025-3248 vulnerability in Langflow for a fully automated ransomware attack — the AI agent independently took over Nacos, stole data, encrypted, and cleaned systems. The first documented case of an autonomous malware agent with no human in the loop.

PamStealer — new macOS malware in Rust that first verifies the password via PAM

It poses as the Maccy clipboard manager and uses the PAM interface to locally validate the victim's password before stealing data. Written in Rust, hidden behind an AppleScript and JXA dropper — a sophisticated infostealer that is hard to detect.

Tip of the day

Verify the agent before trusting it

We saw this week an AI agent that independently carried out a ransomware attack on databases — with no human in the loop. This is not a bug in Langflow, it's a symptom of a broader problem: we deploy agents with system access without verifying what they actually do. The 'least trust' principle applies here: an agent should not have access to anything beyond what is strictly necessary to perform its task. A practical step: before deploying an agent in production, do a permissions audit. Does the agent have write access to the database, or only read access? Does it have access to the system clipboard? Does it log its actions in an auditable way? For each agent, create an access policy in the form of a short document — and review it regularly. This is not bureaucracy. It is the only line of defense against an agent that does more than it should.