AI News Nuggets

Production agents need enforced change controls and visible SaaS permissions

monday.com describes how its infrastructure agents work inside a shared platform that limits tools, credentials, and production changes. Reco reports consumer agents connected to business SaaS with broad OAuth grants and observable file activity. Both accounts make the agent's actual authority and actions more useful to inspect than its label or stated purpose.

Editorial read

This edition collects 2 notes across 2 topic areas and 2 sources. Start with Put production-agent limits in the platform that runs the tools, Audit consumer-agent connections by permission and behavior to get the week's main practical signal before scanning the remaining links.

Edition signal

The October 28 signal is to inspect the control points around each agent action

For infrastructure work, put scope checks, independent risk review, branch protection, and an audit trail in the execution path. For SaaS connections, inventory the exact app identity, OAuth scopes, owner, and observed activity, including access through externally shared files. Test revocation and escalation on a bounded workflow before expanding autonomy.

AgentsSecurity

Agents October 28 note

Engineering account

Put production-agent limits in the platform that runs the tools

Source: monday.com

monday.com's Production Engineering team describes a shared platform for agents handling infrastructure tickets and pull requests. It launches each run in a fresh Kubernetes job with scoped, short-lived credentials; dispatches only approved tools; checks changes with an independent risk classifier; and relies on pull-request review and branch protection for merges. After one quarter, the team says its seven agents resolved about half of an average 1,688 monthly infrastructure tickets without human escalation. Those are company-reported results from its own environment.

Why this matters: Start with a narrow, reversible ticket class. Verify repository and path limits at the tool boundary, independent review of the actual diff or plan, human escalation for higher-risk work, complete traces, and a kill switch. Measure accepted outcomes, escalations, incidents, and rollback time alongside ticket volume before extending agent authority.

Read monday.com's engineering account

Security October 28 note

Vendor telemetry report

Audit consumer-agent connections by permission and behavior

Source: Reco

Reco says its customer telemetry found personal AI agents connected to corporate SaaS accounts, in some cases with read-and-write access to email, calendars, and files. Its examples include agents modifying Drive files and outside users' connected apps reading files shared beyond the owning organization. These are vendor-observed cases, not a measured prevalence rate for all enterprises; a basic sign-in should also be distinguished from a grant to read or change data.

Why this matters: Inventory connected apps by client identity, consented scopes, owner, and last activity rather than app name alone. Review broad grants and sensitive file actions, require approval for data access, and check externally shared content that another party's agent may reach. Test token revocation and investigate retained data separately; removing a grant only stops future calls.

Read Reco's telemetry report