The pattern so far

In August 2026 Anthropic shipped organisation instructions and described them as guidance, not an enforced control. In September the same vendors, and Google and OpenAI with them, moved the other way: settings that block sessions, fail closed on misconfiguration, switch scheduled tasks off and put an expiry on standing approvals. Many data points come from Claude Desktop release notes, which are unusually detailed, so the pattern is read across vendors and not from one changelog.

Trajectory

May2026Jun2026Jul2026Aug2026Sep2026
Status no statusno statusno status Emerging Accelerating

Month by month

  1. Accelerating Reconstructed

    Settings started to stop the product. A disabled Claude Code flag now blocks sessions, Gemini Enterprise's data protection policies inspect and block content in GA, a required Microsoft broker setting fails sign-in instead of falling back, GPT-6 Astra shipped with enterprise access off by default, Gemini blocks URL ingestion inside VPC Service Controls, and Workspace added temporary administrator roles.

  2. Emerging Reconstructed

    Anthropic shipped organisation-wide instructions and called them guidance the model follows, not an enforced control. On the same day OpenAI shipped an Admin plugin that changes workspace settings directly, which put the difference between advice and control on the record.

Before it became a pattern

Early sightings that did not yet pass the entry rule on their own.

← All patterns