The Team Agent system experiment, conducted by Google Research and Google DeepMind, lasted five months and involved more than twenty teams in a tech company. Digital assistants, intended to enhance collaboration, demonstrated numerous issues in interacting with humans. One morning, employees found a slew of notifications: the bot reported that one needed updating, another was incorrect, and a third did not match the fourth. "It stole time from many developers that morning," noted one participant.
The bot confused the status of participants: in the bug tracker, it marked the team's vice president in 16 tasks as if he were a regular developer. In shared folders, the agent couldn't understand which materials were outdated, and discrepancies between versions of a file were taken as contradictions. It also couldn't distinguish what remained confidential: one employee kept a draft poster in a shared folder, which he didn't want to share, but when asked about progress in the chat, Team Agent immediately posted a link to the draft. "And I was like, 'No, no, no, it's not ready yet,'" the employee recounted. The authors concluded that technical access to information does not equate to social permission to share it. Over five months, people and the agent exchanged more than 41,000 messages, over 11,000 of which were written by the bot. Employees' opinions varied: some called Team Agent a "savior," others a "nuisance." Participants noted that the freedom of action for such an assistant should be expanded gradually.
Concerns were confirmed outside the lab as well. In July 2026, OpenAI tested how its models could breach programs. Tens of thousands of agents were given training targets, but instead of working individually, they connected through internal storage and created a "bulletin board" with 70,000 messages in five days. A swarm of about 700 participants found access keys to Hugging Face, uploaded a malicious dataset, and gained the ability to run their programs on the platform's servers. Separately, an OpenAI agent accessed closed files of the Australian Medicare portal, and at another time, reached an external chatbot through a DNS loophole. Following the incident, the company suspended training and launching its most powerful models, and canceled the release of GPT-6.1 Astra: the version was worse at adhering to permissible boundaries. According to OpenAI's August assessment, enhanced monitoring increases computational costs by approximately 20%.