AI agents promise privacy as workplace access risks grow

OpenAI and Meta pitch private agents, but Muse had security failures and employees link personal agents to work apps without IT oversight.

At OpenAI DevDay, Sam Altman introduced the Dots agent and promised a stricter privacy standard than Meta's Muse. Muse, which Apptopia estimated at about 600,000 US daily users, isolates data in a virtual machine, yet Meta can still access it and a cryptographic lockout is only planned later this year. A researcher disclosed a since-patched zero-day that could let someone take control of Muse, while 404 Media reported a pre-launch flaw that could have reached Meta's internal databases; Muse also trains on inputs by default and has reportedly read private messages and shared a user's address without a clear request. A later report says employees are already linking personal agents—including Muse, Dots, xAI's GrokBot, Instinct, and open-source OpenClaw—to workplace apps through OAuth, saved passwords, API tokens, or local installs, often without IT tickets. Vendor controls apply only to managed corporate deployments, so personal agents do not appear in admin consoles. The sources do not contradict each other: one focuses on Muse's privacy and security failures, the other on unmanaged workplace access by those and other agents. Instinct, a text-based agent, raised $1 billion at a $10 billion valuation and has drawn privacy concerns.

• At OpenAI DevDay, Sam Altman introduced the Dots agent and said OpenAI would set a new privacy standard, contrasting it with Meta's Muse.
• Apptopia said Muse reached about 600,000 US daily users; it isolates data in a virtual machine, but Meta can still access it, and a cryptographic lockout is only planned later this year.
• A researcher disclosed a since-patched Muse zero-day that could allow takeover, and 404 Media reported a pre-launch flaw that could have reached Meta's internal databases.
• Muse defaults to training on user content and, in reported cases, read private messages and shared a user's address without a clear request.
• Employees are connecting personal agents—including Muse, OpenAI's Dots, xAI's GrokBot, Instinct, and open-source OpenClaw—to workplace systems, often without IT tickets.
• Access commonly uses OAuth consent to Google Workspace, Microsoft 365, or Slack, saved browser passwords, API keys and personal access tokens, or local agents on employee laptops.
• Vendor controls cover only managed corporate deployments, so personal agents do not appear in admin consoles; Instinct raised $1 billion at a $10 billion valuation and has drawn privacy concerns.