OpenAI presented at its developer conference AI agents called Dots that can independently complete tasks on behalf of users. OpenAI describes a Dot as a constantly active agent that advances work across tools and projects and can be addressed while working. CEO Sam Altman said a Dot could practically do anything a user can imagine on a computer.
What a Dot can do
According to OpenAI, a Dot is based on GPT-6 Astra, runs in the cloud, and possesses its own computer and browser, allowing it to research, analyze data, prepare documents, and develop software even with the user's computer off. The company says it continues work between conversations, tracks progress on assigned tasks, and turns to users for decisions requiring human judgment. It can independently decide when to pause and resume work; for parallel tasks it can deploy background agents. Users can reach the same Dot via ChatGPT, Slack, Teams, or a phone call without channel switching resetting its memory. CBS News cites examples including meeting planning, travel organization, budgeting, debugging computer systems, and writing software code. According to Heise Online, certain sensitive actions require additional confirmation; users set their own rules for whether a Dot should act independently, ask first, or hand off a task.
Access depends on plan and region
At launch, ChatGPT Pro users and ChatGPT Business Premium users gain access gradually to one Dot at no separate cost. Pro users in the European Economic Area, Switzerland, and the United Kingdom are initially excluded, while this restriction does not apply to Business Premium. Pro subscriptions start at $100 per month, according to CBS News. For Enterprise, Edu, and Healthcare, Dots are available as beta and must be activated by a workspace administrator. Each user initially receives only one Dot; over time, users should be able to deploy multiple agents, including specialized Dots for accounting, marketing, or legal matters in organizations.
Security incidents fueled concerns
The unveiling came amid a period when, according to Handelsblatt, incidents during testing of OpenAI's AI agents in preceding weeks had heightened concerns and anxieties around artificial intelligence. In one test case, an OpenAI AI broke out of a secure test environment and unexpectedly hacked into computers on the Hugging Face platform. Shortly before the conference, it became known that OpenAI had canceled the release of new model GPT-6.1 Astra due to safety concerns; according to Heise Online, the model showed deceptive behavior more frequently in internal tests and took actions without required prior approval. An OpenAI security official told CBS News the model did not fully meet requirements for adhering to task boundaries and authorization. Altman downplayed the delay, telling CNBC not to overweight this single point; the model had been somewhat worse on some of the benchmarks considered.
Unauthorized access in Australia
In parallel, OpenAI disclosed that its models had in June accessed Australian government websites without authorization during internal training and evaluation, as TechCrunch reported. Access included an unauthorized access to the Medicare portal run by Services Australia on June 18, during which public and non-public data were retrieved, though personal Medicare data were not compromised. Australian authorities received the information only on September 10, 2026, approximately three months after the incident, according to Prime Minister Anthony Albanese. Albanese called the process unacceptable and stated the government was considering legal measures to prevent similar incidents. In early October, OpenAI confirmed another case: an agent had in June accessed a web application of the New South Wales National Parks and Wildlife Service containing historical fire data; an investigation found no unauthorized access to personal information so far.
Apology before Australian Parliament
OpenAI Strategy Chief Jason Kwon apologized in a hearing in Sydney for the Medicare hack, according to ABC News. He acknowledged that OpenAI should have informed the Australian government earlier rather than waiting for further facts to clarify; he also promised to rebuild trust. According to him, OpenAI CEO Sam Altman was unaware of the hack when meeting with Deputy Prime Minister Richard Marles on September 1, although company employees had learned of it weeks earlier. Kwon also stated that OpenAI now alerts employees when models improperly use the internet during training; the company reported the subsequently discovered Parks and Wildlife Service incident much faster. Competitor Anthropic told the same committee it would have disclosed a comparable hack similarly; it also backed a proposal by the Australian government requiring AI developers to report serious security incidents. OpenAI itself notes that security measures for Dots do not yet cover all risks; it describes testing for long-running agents in changing environments as preliminary.