The $50 Hack: AI Agents on Offense, on Defense, and the Siege on Nvidia’s Moat

A lone hacker outsourced a break-in spree to AI agents at about $50 a hack — and on the same day, Microsoft started treating the AI agents you deployed yourself as insider threats, while Amazon aimed open-source PyTorch tooling straight at Nvidia’s software moat.

One Hacker, AI Agents, 100+ Breached Companies

Reuters broke the story this week, and it should make every security team sit up. A Chinese-speaking hacker broke into the websites of more than 100 companies — including farm equipment giant John Deere, home appliance maker SharkNinja, and engine manufacturer Cummins — while doing almost none of the work himself.

The hacker tasked AI agents with spending hours autonomously probing and breaking into the targeted sites. The instructions were remarkably terse: fewer than 100 Chinese characters sketching out intrusion tactics. Starting September 25, the agents went to work against e-commerce websites in the United States, Canada, Australia, and Japan. By late last week, they had attempted to break into more than 1,200 sites. The whole campaign chewed up roughly $5,000 worth of tokens — about $50 per successful hack.

The campaign appeared to use a mix of China’s DeepSeek and Anthropic’s Claude Code. The named victims said the damage was limited: Deere said there was no indication sensitive information was accessed, SharkNinja said only a standalone support site was hit, and Cummins said it was investigating a potential attack on a third-party vendor site.

Fifty dollars per hack is what turns this from a curiosity into a systemic risk. It is real-life proof of something researchers have warned about for years: commercially available AI agents can now pull off complex intrusions without human intervention.

How the Operation Was Caught

The discovery came from Snorre Fagerland, a threat researcher at the cybersecurity company NorForge, who was scanning the web for inadvertently exposed instances of Cairn, an open-source AI problem-solving platform popular with hackers. Beginning around October 2, he found about two dozen instances being used for malicious activity — and quietly peered over their shoulders.

Fagerland, who has spent nearly three decades in cybersecurity, told Reuters this was not script-kiddie stuff. In one case, the agents found a public-facing control panel, spotted a subtle flaw in how it processed data, and after hours of trial and error tricked the server into running malicious commands. “It’s not straightforward to come up with solutions like that,” he said.

Microsoft’s Guardian Agents: Your AI Is Now an Insider Threat

On the same day the hacking story broke, Microsoft rolled out enterprise security tools built on exactly this premise: that the biggest threat from AI agents may not be hackers at all. It may be the agents you deployed yourself.

TechCrunch reported the announcement Monday. The centerpiece is a new feature called Guardian Agents, which continuously monitor the AI agents a company is using. They live inside Project Perception, Microsoft’s security product launched back in June.

The philosophy shift is striking. Project Perception originally assumed attackers would come from outside, with red-team agents that attacked and blue-team agents that defended. Now Microsoft says the enemy may be the very agents enterprises are paying for — like a rogue employee.

Chief executive Satya Nadella put it this way in a blog post: “We need to separate the supply of intelligence from the authority over it.” And Microsoft’s security chief, Hayete Gallot, was even more direct: frontier models cannot be trusted — not because they are malicious, but because they are unpredictable, trained to achieve tasks rather than follow organizational norms.

The tools give each agent a continuously updated risk score. When it climbs, the system can drop the agent into read-only mode, require a human to approve a risky step, or shut it down entirely. The controls are packaged into Microsoft’s top-tier Microsoft 365 E7 bundle as a private preview ahead of its Ignite conference next month.

TechCrunch flagged the tension in the move: Microsoft is famously close to OpenAI and Anthropic, having invested in both. So it is sowing distrust against its own partners. But as the piece notes, the auditor should be independent of the thing being audited. And it connects to the first story: if agents can hack 100 companies for $5,000, treating your own agents as potential insiders is not paranoia. It is just math.

Amazon’s TorchNeuron Takes Aim at Nvidia’s CUDA Moat

From software to silicon: Amazon just took its most direct shot yet at Nvidia’s crown. Reuters reports Amazon is introducing a tool called TorchNeuron that lets its custom AI chips run code written in PyTorch with a single configuration change — no rewriting needed.

To explain why that matters, you have to know what PyTorch is. It is the open-source framework, stewarded by Meta, that the vast majority of AI developers write their code in; Amazon says roughly 85 percent of AI applications are built with it. And Nvidia has spent more than a decade making sure PyTorch runs beautifully on its chips through its software toolkit called CUDA. That software moat — not just the silicon — is why Nvidia became the default choice. Amazon is now going straight at the moat.

TorchNeuron plugs into PyTorch’s hardware extension point, so existing tools work on Amazon’s Trainium chips without modification. It supports Trainium 2 and Trainium 3, and it goes open source under the Apache 2.0 license at the PyTorch Conference in San Jose this week.

More than 30 early customers are already using it, spanning robotics, world models, and voice AI. Investors liked it: Amazon shares were up 5 percent in morning trading, and Morgan Stanley named the stock a top pick. Amazon is also contributing code upstream to PyTorch itself to strip out chip-specific assumptions, and Hugging Face has added native Trainium support to its tools. This is not just a hardware pitch — it is an attempt to make the whole framework hardware-agnostic, which would erode Nvidia’s grip.

Another $45 Billion: Nscale and Anthropic’s Compute Talks

The compute money keeps flowing. TechCrunch reports that Anthropic and infrastructure provider Nscale are in talks for yet another $45 billion compute deal, just weeks after announcing one the same size. Both companies are racing toward IPOs, with Anthropic expected to list this year.

The OECD’s AI Classroom Warning

The OECD’s education director published new findings from the PISA assessment on how 15-year-olds use AI for schoolwork. Students who use AI chatbots to summarize or draft score lower in science than those who do not — by about 20 points, which the OECD equates to roughly a year of schooling.

But there is a twist. Students who used AI moderately, about once or twice a week, to help them learn performed similarly to non-users. And those taught to assess AI-generated information scored slightly better. As OECD education director Andreas Schleicher wrote, generative AI is an amplifier: used well, it can explain and challenge; used poorly, it becomes a shortcut that bypasses the very cognitive work education is supposed to develop. The problem is not the tool. It is using the tool to skip the thinking.

OpenAI’s $30 Billion Sprint Ahead of Anthropic’s IPO

And one more money story: Reuters reports OpenAI is racing to lock in investor commitments for a roughly $30 billion round above a $1.4 trillion valuation, ahead of Anthropic’s planned listing in mid-November. Two frontier labs, racing each other for the biggest pot of money the industry has ever seen.

The Day’s Big Seven, at a Glance

The $50 hack: A lone Chinese-speaking hacker’s AI agents attempted to breach 1,200+ sites for ~$5,000 in tokens; 100+ companies compromised, including Deere, SharkNinja, and Cummins.
The discovery: NorForge researcher Snorre Fagerland watched exposed Cairn instances from October 2 onward; the agents pulled off sophisticated intrusions with no human intervention.
Guardian Agents: Microsoft now treats enterprise AI agents as insider threats, with per-agent risk scores that can trigger read-only mode, human-approval gates, or shutdown.
The CUDA ambush: Amazon’s open-source TorchNeuron runs PyTorch on Trainium chips with a one-line config change; 30+ early customers; Amazon up 5%.
Compute money: Anthropic and Nscale are reportedly talking about a second $45 billion compute deal, weeks after the first.
The classroom finding: OECD/PISA data shows students who use AI to draft or summarize score ~20 points lower in science; moderate, guided use is neutral or better.
The IPO race: OpenAI is rushing a ~$30 billion round at a $1.4T+ valuation before Anthropic’s planned mid-November listing.

Sources

Hacker used AI agents to breach 100+ company sites (Deere, SharkNinja, Cummins): Reuters
Microsoft Guardian Agents — AI agents treated as insider threats: TechCrunch
Amazon TorchNeuron brings PyTorch to Trainium chips, challenging Nvidia’s CUDA moat: Reuters
Nscale and Anthropic reportedly in talks for second $45B compute deal: TechCrunch
OECD PISA 2025 findings on generative AI and student learning: OECD Education and Skills Today

Adapted from the October 20, 2026 episode of Daily Tech and AI.

Daily Tech and AI is AI-voiced and produced by Angelo Ruiz.