OpenAI: Agents, Security, Privacy

OpenAI is pushing deeper into workplace automation and trust tools, adding shared agents in ChatGPT while briefing governments on cyber defense software. (openai.com) OpenAI said on April 22 that workspace agents are rolling out in research preview to ChatGPT Business, Enterprise, and Edu customers. The company said teams can build one agent, share it across a workspace, and run it inside ChatGPT or Slack. (openai.com) The new agents are powered by Codex, run in the cloud, and can connect to tools including Google Drive, Google Calendar, Slack, and Microsoft SharePoint. OpenAI’s help and release notes say eligible workspaces can also schedule agents, add files and skills, and connect custom Model Context Protocol servers. (openai.com, help.openai.com) A workspace agent is OpenAI’s latest version of a custom GPT for office work: instead of answering one prompt, it can carry out a repeatable, multi-step task across company systems. OpenAI’s Academy materials frame the product around shared workflows, standard handoffs, and outputs that teams can reuse over time. (openai.com) At the same time, OpenAI has been pitching a separate cybersecurity product to governments. Axios reported, and Reuters matched the report on April 22, that OpenAI briefed U.S. federal agencies, state governments, and Five Eyes allies on the product’s capabilities over the past week. (axios.com, reuters.com) Reuters said it could not independently verify the Axios report, and OpenAI did not immediately respond to Reuters’ request for comment. Reuters also noted that cyber tools have become a competition point for OpenAI and Anthropic as governments and companies weigh both the defensive uses and the misuse risks of advanced models. (reuters.com) OpenAI also released an open-source privacy filter on April 22 for detecting and masking personally identifiable information in text. The company said the model is small enough to run locally, so unfiltered data can stay on a device instead of being sent to a server first. (openai.com) In its model card, OpenAI described Privacy Filter as a bidirectional token-classification model with span decoding for PII detection and redaction in unstructured text. The company said it is built for high-throughput privacy workflows and published the model under an Apache 2.0 license. (openai.com, openai.com) Taken together, the April 22 releases show OpenAI tying three enterprise demands into one week of launches: automating work inside company software, selling security capabilities to public-sector buyers, and reducing privacy exposure before data moves through an artificial intelligence pipeline. OpenAI’s newsroom listed all three announcements within days of each other. (openai.com) The next test is adoption inside large organizations, where buyers want automation that can touch internal systems without losing control of permissions, security, or sensitive data. OpenAI’s product pages now pitch workspace agents with governance controls, while its privacy and cyber moves aim at the same trust problem from different sides. (openai.com, openai.com)