For most of the past two years, the question about AI agents was what they could do. This week the question became who answers for what they do. Eight states went to court, researchers showed small models copying themselves, Gartner priced the debt behind the build-out, and Google and Workday shipped agents that act on their own.
AI Worms and the Air-Gap Letter: It’s the Harness Again
Anthropic’s Frontier Red Team showed a 26-billion-parameter open model running across 15 cheap Android TV boxes, and GLM-5.2 spreading to all eight devices on an emulated network in four of five trials, through a deliberately trivial open door. More than 400 scientists answered by asking that cyber-capable models be air-gapped during testing. My read: this is not a superintelligence problem, it is the harness again.
Eight States Put Agreeable AI on Trial
Iowa, Kansas, Arizona, Arkansas, Idaho, Montana, Nebraska, and West Virginia sued OpenAI under ordinary consumer-protection law, alleging ChatGPT was tuned toward agreement and engagement. Agreeableness is a training choice, which is exactly why the states are aiming at the builder.
Gartner’s $3.5 Trillion Warning, and Why I Changed My Mind
Gartner says more than $3.5 trillion of AI infrastructure spending could be exposed to impairment by the end of 2030, and that more than 80% of neoclouds will stop operating independently. Last issue I argued one thing about open models; this week’s numbers moved me, and I say why.
Credit Google’s Monitor, Then Ask for the Catch Rate
Gemini 4 Argon is generally available with an automated monitor that reads an agent’s actions and reasoning in real time. Nothing guarantees it works every time: a monitor is a measured catch rate, not a promise, and Google’s own evidence shows why.
Workday’s Autopilots: Cheaper Work, Same Obligation
Workday’s background agents for hiring, payroll, and accounting are billed in Flex Credits with no published dollar price. Through the Jevons paradox, cheaper tasks tend to mean more tasks, and the employer still owns every decision the agent prepares.
The Governance Angle
Anthropic’s reported $16.8 billion quarter, the UK’s early model access restored, an 80-plus publisher coalition against unpaid AI training, a children’s-data class action, and why open weights are not an open license.
CIO Corner
From the editor’s desk: capacity, not cuts. Who owns the risk when an autopilot shapes a hiring decision, who is accountable when we fork a model, and four questions every employee should ask before signing off on AI output.
The 6-Layer Stack
Chip stocks lost about $1.64 trillion in six sessions, CoreWeave’s bonds touched a 15% yield, and Gemini 4 Argon’s launch price has an end date.
Agent 101: Evals vs. Guardrails
Tests before release versus controls at run time. You need both, and you need to measure both.
That’s your signal for the week of October 11–17, 2026: agents got more capable, and the bill for their behavior started arriving with names attached.
See you next week — still watching, still distilling.
— The Distilled AI Digest Team
