Today's Headlines

  • Exposed control panels show AI agents running break-ins, with 1,100+ attack projects aimed mostly at the US, Japan and Australia
  • Amazon unveils TorchNeuron to run PyTorch code on Trainium without a rewrite, open-sourcing it on October 20
  • Inception releases Mercury 3, a diffusion LLM running at 1,068 tokens per second with a 1M-token context
  • AI chipmaker Tenstorrent raises $1.5 billion at a $6.2 billion post-money valuation

The day's thread is how much work is now handed to AI. Security firm NorForge found control panels for an AI-driven hacking system sitting open on the internet, revealing how attackers point autonomous agents at their targets. On the building side, Amazon unveiled a tool that lets developers run the PyTorch code they already have on its Trainium chips.

Today's Top Three

Exposed control panels show AI agents running break-ins, with 1,100+ attack projects aimed mostly at the US, Japan and Australia

Security firm NorForge reported on October 19 that many control panels for Cairn, a system that hands intrusions to AI agents, were reachable on the open web with no login.

Cairn is open-source. An operator sets a starting point and a goal, and a group of AI agents works toward it, posting what they find and what they plan to try next to a shared board. NorForge identified more than 20 exposed panels it considers likely to be malicious.

The most active group works in Chinese and had over 1,100 attack projects across 14 panels. Its targets are mostly organizations in the US, Japan and Australia, and the aim appears to be lasting access to e-commerce and payment systems. One typical task listed shopping sites in four countries, already checked by script to accept orders, and told the agents to find a way to run code without logging in.

The group's agent worker is named "claudecode_deepseek-v4.1-flash", which NorForge reads as a possible Claude Code harness running a DeepSeek model. NorForge estimates an average cost of around $5 per task and a success rate of around 10%.

Reuters reported that the victims included John Deere, SharkNinja and Cummins. NorForge says it is working with Reuters to notify victims and is pressing hosts to take the panels down.

According to The Verge, SharkNinja said the breach hit a standalone product-support site and has been contained. Cummins said it is investigating a possible attack on a website run by one of its vendors.

Amazon unveils TorchNeuron to run PyTorch code on Trainium without a rewrite, open-sourcing it on October 20

Amazon announced TorchNeuron on October 19, a tool that runs AI code written in PyTorch natively on its Trainium chips.

Switching chips takes a single configuration change. TorchNeuron plugs into the standard mechanism PyTorch already provides for registering new hardware, so the tools developers use to test, debug and scale their models keep working. Amazon will release it under an open-source license at PyTorch Conference North America on October 20.

Amazon says 85% of AI researchers and developers write their code in PyTorch. Until now, moving to a new chip meant rewriting and revalidating code that had taken months to perfect. Amazon is also working with the PyTorch community on upstream changes that remove code tying the framework to specific chips.

Colin Brace, VP of Technology Software at AWS Annapurna Labs, the unit that makes Amazon's custom chips, told TechCrunch that Trainium's value lies in its price-performance and in serving both training and inference on one platform.

Inception releases Mercury 3, a diffusion LLM running at 1,068 tokens per second with a 1M-token context

Inception released Mercury 3 on October 19, a diffusion-based language model that refines a whole passage at once instead of writing it word by word.

According to the company, Mercury 3 generates 1,068 tokens per second on NVIDIA GPUs, and its context window has grown to 1 million tokens from 260,000 in Mercury 2.5. It has more than 40 times the total parameters of the first Mercury model.

Inception says Mercury 3 matches or beats speed-focused models such as GPT-6 Luna and Claude Haiku 5.5 while running four to eight times faster. It costs $0.20 per million input tokens and $0.70 per million output tokens, with a 50% discount at launch.

The model is available through Inception's own API and through OpenRouter, Baseten and Vercel, and it is compatible with the OpenAI API. Early-access partners are using it where latency matters most, including voice and chat support, tutoring and web-browsing agents.

Other Developments

Models

Policy

  • OpenCode, the maker of an open-source AI coding agent, said over the weekend that attackers broke into its online platform, OpenCode Console, affecting 1.4% of its users. TechCrunch reports that users saw charges on their OpenAI, Anthropic and Google AI accounts from stored API keys, in the hundreds of dollars on individual accounts and above $1,000 for a few. OpenCode says refunds for unauthorized charges are underway. Hackers hit AI coding tool OpenCode, running up charges on users' AI accounts (TechCrunch)
  • A federal magistrate judge ordered on October 19 that the former OpenAI engineer charged with planting recording devices in OpenAI's conference rooms stay in custody until trial. Ars Technica reports that the government's detention motion says the FBI found he had opened a folder called "Saudi Gov Docs" on his iPhone before his arrest. Accused OpenAI spy had folder called "Saudi Gov Docs," FBI says (Ars Technica)

Business

Source: Selected by the editors from the AI news inbox (collected October 20, 2026: 38 items, 4 primary and 34 secondary).