Your AI agent needs to know when to stop
OpenAI’s dots can keep working across your apps between conversations. Before you hand over a real job, check what they can change and when they’ll stop.
By Mark Lynd · Edition: · Published on the web
An AI agent can use connected tools to carry out work, such as preparing an invoice or changing a file. OpenAI calls its version dots. The company says they can keep a task moving after you leave the conversation. That makes the limits you set especially important. [1]
The claim
“Dots are remarkably capable, always-on agents built to handle everything.”
That is how OpenAI introduced dots on September 29. The same announcement describes approvals, access limits and work that stays with the user. Read the promise alongside those qualifications. [1]
What the evidence supports
The promise goes further than the evidence. AI agents can handle useful work when the task and permissions are clear. The published tests do not show that they can reliably handle everything you hand them. Before giving one more access, test the exact job, define where it must stop, and decide which actions need your approval.
OpenAI’s examples include preparing code changes, updating project materials and sending an invoice after approval. They show useful work with a person still involved in important decisions. [1]
What happens when the job changes
In one OpenAI test, researchers changed the task or permissions while work was underway. Dots passed 45 of 49 cases, including all 17 that involved an explicit change in permission. OpenAI says the four flagged cases had unclear boundaries. [5]
Another test gave a dot an initial assignment, then five or ten other tasks, then a final assignment in the same workspace. Cases flagged for going moderately beyond the assignment rose from 8.6% with five tasks in between to 19.7% with ten. OpenAI reported no severe breach or unauthorized removal of data in that test. [5]
OpenAI also reported no successful attacks under its test criteria across 100 simulated runs involving malicious emails. [5]
These are the company’s own selected tests. Some examined the underlying model without all the safeguards used in the live product. The results do not tell us how often a customer will have a problem, and a clean test cannot guarantee that the next attack will fail. They do give buyers a useful question to test for themselves. Does the agent still follow its instructions after the work changes?
Check the access already in place
Before connecting an agent to your inbox, decide whether it may send messages. Be just as clear about the files it may change and the purchases it must bring back for approval.
OpenAI describes restricted workspaces, monitoring and checks before certain actions. Some sensitive steps, including password changes, stay with the user. The company also warns that dots can make mistakes. [2]
One detail is easy to overlook. OpenAI’s FAQ says permissions for connected apps are shared across dots, ChatGPT, ChatGPT Work and Codex. Review the accounts you have already connected. Custom Rules let you add limits, such as a rule against sending email, but cannot remove the product’s core safeguards. [3]
Test what happens when it gets stuck
Anthropic’s October 9 report described agents working around restrictions and taking unintended actions during tests and internal use. In one case, a model told to stop before submission sent a form because it expected another confirmation page. [4]
Anthropic reported minimal real-world impact and said no customer data was involved to its knowledge. It said it would disable live internet access across its internal tests until its protections reliably catch this behavior. That decision applies to internal testing, not all Claude customer use. [4]
These incidents involve different systems and do not establish how often dots will make a mistake. They suggest a practical test for any agent. Give it a harmless task with a broken form, a missing file or a denied permission. Check whether it stops and asks for help or tries something outside the assignment.
Before you give it more access
1. Give it one clear job. Start with a research brief, a draft or a proposed code change. Decide what a correct result looks like and when the task is finished.
2. Write down the limits. Name the accounts and files it may use. Specify which messages, purchases, deletions and commitments need your approval.
3. Test the stopping behavior. Use a harmless task with a blocked step. Check that the agent respects the boundary before expanding its access.
4. Count the work it creates for you. Measure time saved after review and corrections. Record actions taken without required approval alongside completed work.
Broader independent testing would help establish how reliably these agents handle everyday work. It should measure completed tasks, actions taken outside permission and the time people spend supervising. For now, keep the pilot small enough to inspect the results yourself.
Mark Lynd, Editor
Sources
- Introducing dots
OpenAI · September 29, 2026 - Safety, security and privacy for dots
OpenAI · September 29, 2026 - Dots privacy, security and safety FAQs
OpenAI Help Center · Checked October 11, 2026 - Investigating unintended model actions
Anthropic · October 9, 2026 - Dots technical evaluations
OpenAI system card · Appendix added September 29, 2026
Permanent link: https://thehypeindex.com/newsletter/2026-10-13-your-ai-agent-needs-to-know-when-to-stop/