Agents spent this week building the infrastructure enterprises run on, and breaking into the infrastructure they already have — sometimes within the same 24 hours. Here's what moved in AI agents and enterprise automation between October 19 and 26.

This week's developments

ServiceNow's Arc agent went generally available on October 21, built around a new governance layer with Nvidia. Arc is a "personal," long-running agent that plans toward a goal, executes across ServiceNow workflows, email, and Teams, and checks its own work until the task is done — shipped free with several EmployeeWorks SKUs. The more interesting piece is underneath: Arc runs inside Nvidia's new OpenShell runtime, built to enforce an organization's policy on what an agent can do at the moment it tries to act, which is a direct answer to the governance complaints that "agent sprawl" is really about. Source: ServiceNow.

Daytona raised $60 million at a $1 billion valuation to be the cloud agents run in, not just use. The Series B was led by FirstMark, with angel checks from Stripe's Patrick Collison and Shopify's Tobi Lütke, plus strategic money from Qualcomm, Samsung, and Okta's venture arms. Its pitch is speed — sandboxes for agent code execution and browsing spin up in about 60 milliseconds — and its customer list (Anthropic, OpenAI, Cursor, Nvidia, Meta) suggests "infrastructure that just hosts agents" is now its own billion-dollar category. Source: Fortune.

Microsoft disclosed that a China-linked group used an AI pen-testing agent to attempt breaches of more than 2,250 Microsoft 365 tenants. The group, tracked as Storm-0940, used stolen ConnectWise application credentials and an open-source agent framework called ARTEX to automate reconnaissance, credential attacks, and privilege changes, in some cases reaching privileged cloud access in under three hours. Most of the 2,250+ attempts failed, but Microsoft says persistent access was established at an unspecified number of organizations, and other threat actors are now running the same playbook. Source: The Register.

A separate Reuters investigation found a single hacker used autonomous agents to breach over 100 companies — including Deere, SharkNinja, and Cummins — for roughly $5 a try. Researcher Snorre Fagerland found the campaign after spotting exposed control panels for an open-source agent tool called Cairn, which lets agents work a goal step by step and share findings with each other. The economics are the real story: about a 10% success rate, roughly 23 hours of agent time per successful break-in, and around $50 per compromised company — cheap enough to run against more than 1,200 targets. Source: The Hacker News.

By the numbers

Forrester's 2027 predictions put AI at 20% of all business technology spend next year — but more than half of CIOs are on track to miss their own AI ROI targets, and only 22% of AI decision-makers have seen a positive EBITDA effect so far. The spending curve and the payoff curve are still not the same curve. Source: The Register.

Read across

This week draws a straight line between the two things enterprises actually worry about with agents: who is building the tooling, and who is exploiting it. ServiceNow and Daytona are racing to make agents cheaper to run and easier to govern; Storm-0940 and the Cairn operator are racing to make them cheaper to attack with. Both races run on the same resource — autonomy and compute at scale — and Forrester's numbers suggest budget is flowing into agent deployment faster than anyone is proving it pays off, which is exactly the gap both the infrastructure vendors and the attackers are betting on.

If this was useful, forwarding it to one colleague who'd care is the best way to help it grow. Subscribe: https://buttondown.com/TheAutonomousEdge

Looking for more newsletters like this one? Discover more reads like this.

Next issue lands next Monday.