October 18–24, 2026 · Issue #41
Agents worked both sides this week. One launched to run your day for you; others let a single attacker break into more than 100 companies for about $50 a hit.
1. Claude Duet and the Always-On Agent: You Will Pick One
Anthropic opened a research preview of Claude Duet, an agent that comes to you instead of waiting for a prompt. It joins Grok Bot, Muse, Dots and the Gemini agent, and the market is splitting: organizations will pick one, not mix and match.
2. $50 a Break-In: Ask Why, Not How
Reuters reported that a Chinese-speaking hacker used AI agents to breach more than 100 companies, including Deere, SharkNinja and Cummins. At roughly $50 per successful hack, every company is worth trying. Assume you are being tested cheaply, at volume.
3. Gartner Maps 2027, and the Harness Is the Door
Gartner’s top strategic technology trends for 2027 predict that by 2028, 25% of enterprise breaches will trace to AI agent abuse. That number assumes nothing changes. The agent harness is where it gets bent down, if the investment shows up.
4. The Bill Is the Brake: Splunk’s 55%
In Splunk’s survey of 2,050 IT operations and engineering professionals, 55% say they cannot control AI costs. Note the source: Splunk sells the tools it recommends. Unpredictable bills are now the reason leaders pause AI investment.
5. LiteLLM, Again: Vendors Can Sell the Hardening, Not the Accountability
The open-source AI gateway was compromised for the second time this year, this time through its release pipeline. Buy the hardening from a vendor or build it yourself, but the accountability stays with you.
The Governance Angle
Homeland Security’s DeepSeek order reaches products built on the model, Microsoft’s AI diagnosis result comes with fine print, and Anthropic sets aside shares for public benefit ahead of its IPO.
CIO Corner
Start with the sponsor’s number. Where to begin on the agent road, and why every pilot needs a spend cap and a named cost owner.
The 6-Layer Stack
Amazon and Google both open-sourced PyTorch paths onto their own chips, a direct hit on Nvidia’s software moat.
Agent 101
Pencil versus pen: why the agent actions you cannot undo are the ones that need a human.
The signal this week: agents are now cheap enough to work for anyone, on either side, so the controls you put around them decide whose side they work for.
See you next week — still watching, still distilling.
— The Distilled AI Digest Team
