Breaking

OpenAI launches dots as always-on agents while safety scrutiny shifts to actions

September 30, 2026 at 05:11 EDT

AI Agents
Security
Foundation Models

OpenAI launched dots on September 29, 2026, bringing always-on AI agents into ChatGPT for paid users while the industry is under new pressure to show that autonomous systems can stay inside the lines. The product turns the next safety question from what a chatbot says to what an assistant does when it has tools, memory and access to user accounts.

Sep. 29, 2026 · OpenAI

dots puts agent safety on stage

OpenAI's always-on assistant can keep working across tools and accounts. The launch turns the safety question from what an AI says to what it does.

50,000

simulated emails in a bulk prompt-injection test

16,600

attack emails inside that test

scored attack successes reported

91.8%

pass rate on changing permission instructions

Longer context, higher flag rate

When intervening tasks doubled from five to ten, observed flags rose from 8.6% to 19.7%.

8.6%

5 intervening tasks

19.7%

10 intervening tasks

How the handoff is supposed to work

Set a goal

Give dots an ongoing responsibility.

→

Connect tools

Choose apps, memory and access.

→

Review actions

Use Custom Rules and Auto-review.

Why teams may try it

Persistent follow-up could help with RFPs, customer feedback, API migrations, inbox triage and work that falls between meetings.

Why trust is fragile

The value depends on users understanding access, approvals, memory, logs and when a dot can act without another prompt.

The real test is control

dots launches with a control stack, but long-running agents face changing context, ambiguous authorization and account-level consequences.

Custom RulesAuto-reviewHuman approvalActivity review

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.