OpenAI Launches ChatGPT Agent for Enterprise Users
OpenAI has released ChatGPT Agent, a system that merges three existing capabilities – web interaction, research synthesis and conversational intelligence – to provide a true agentic AI experience for ChatGPT subscribers.
According to OpenAI’s release notes, the agent can navigate websites, filter search results, execute code and produce editable documents including spreadsheets and presentations. Users can request actions such as calendar briefings based on recent news, ingredient purchasing for meal planning or competitor analysis with accompanying slide decks.
How ChatGPT Agent builds on OpenAI’s existing tools
The release represents an evolution of OpenAI’s Operator and Deep Research capabilities. Operator could interact with websites through clicking and typing, while deep research handled information analysis and summarisation. Each tool operated within specific parameters that limited cross-functionality.
The agent system addresses these limitations by combining both approaches within a single interface. Users can transition from conversation to action requests within the same chat session. The system employs multiple tools including a visual browser for graphical interfaces, a text-based browser for reasoning tasks, terminal access and direct API connections.
ChatGPT connectors enable integration with applications including Gmail and GitHub, allowing the agent to access relevant information for responses. Users can log into websites through browser takeover mode, expanding the agent's research and task execution capabilities.
The “virtual computer”, as OpenAI describes it, maintains context across different tools, enabling the agent to open pages in various browsers, download files, manipulate them through terminal commands and display results in the visual browser.
Kevin Weil, OpenAI’s Chief Product Officer, has previously framed 2025 as a transition year for AI capabilities. “I think 2025 is the year that we go from ChatGPT being this super smart thing that can answer any question you ask to ChatGPT doing things in the real world for you,” he told Axios in January.
“You can just have it go off and do a bunch of work,” Kevin said at the Human[X] conference in March 2025. “Then it’s not just how efficient can you make yourself, it’s also how many agents can you have [going] off solving problems for you.”
OpenAI addresses security concerns with ChatGPT Agent
The agent’s web interaction capabilities introduce new risk factors, particularly regarding data handling through connectors and logged-in websites. OpenAI has implemented safeguards addressing prompt injection attacks, where malicious instructions embedded in web content could manipulate agent behaviour.
- Performance leap: ChatGPT agent achieved 41.6% accuracy on Humanity's Last Exam expert-level questions, rising to 44.4% with parallel attempts
- Enterprise reach: OpenAI serves over 400 million weekly ChatGPT users, with 3 million developers using its API and 2 million business users on enterprise products
- Efficiency gains: Agent can complete spreadsheet tasks with 45.5% accuracy compared to Microsoft's Copilot in Excel at 20.0%, while matching or exceeding human performance in approximately half of complex knowledge work scenarios
The company trained the agent to identify and resist prompt injections while implementing monitoring systems for rapid attack detection and response. User confirmation requirements for consequential actions provide additional protection against unauthorised activities.
The system includes explicit user confirmation for actions with real-world consequences, active supervision for critical tasks such as email sending and proactive refusal training for high-risk activities including bank transfers.
Privacy controls enable users to delete browsing data and log out of active sessions with a single click. During browser takeover mode, user inputs remain private with no data collection or storage of sensitive information such as passwords.
OpenAI has classified the agent as High Biological and Chemical capabilities under its Preparedness Framework, activating associated safeguards despite lacking definitive evidence of severe biological harm potential. The company describes this as the most comprehensive safety stack implemented to date, featuring threat modelling, dual-use refusal training, always-on classifiers, reasoning monitors and enforcement pipelines.
OpenAI says it works with biosecurity experts, safety institutes and academic researchers to shape threat models and policies, with biology-trained reviewers validating evaluation data while domain-expert red teamers tested safeguards in realistic scenarios.
OpenAI CEO Sam Altman expects workplace transformation to accelerate this year, writing in his January blog post that OpenAI believes 2025 may see the first AI agents join the workforce and materially change company output. “We believe that, in 2025, we may see the first AI agents ‘join the workforce’ and materially change the output of companies,” he wrote.
