As reported by BleepingComputer, OpenAI is testing a new "Writing Style" feature for ChatGPT that connects to users' personal apps — including Slack, Gmail, Google Drive, and Notion — to learn and replicate their writing voice. While the consumer framing is about personalization and convenience, the security implications deserve scrutiny before this reaches general availability.
Why This Matters from a Security Lens
This feature represents a meaningful escalation in how AI assistants access user data. Previous personalization features — like Anthropic's Claude Styles — relied on users actively pasting or uploading samples. OpenAI's approach instead pulls directly from connected services via OAuth integrations, which means ChatGPT is now granted read access to potentially vast corpora of personal and professional communications.
The distinction matters. When a user pastes a sample, they maintain conscious control over what's shared. When an integration reads from Gmail or Slack, the scope of accessible data is governed by the OAuth scopes granted — which users rarely read carefully. A "read your emails" permission may expose years of correspondence, including sensitive attachments, proprietary information, and confidential communications.
Enterprise Risk Amplification
If this feature rolls out broadly, expect it to create tension in enterprise environments. Employees who connect work Google Drive or Slack accounts to ChatGPT for writing assistance may inadvertently route corporate data through OpenAI's infrastructure. Key concerns:
- Data exfiltration surface: OAuth tokens tied to ChatGPT become a new vector for sensitive data to leave controlled environments.
- Scope creep: Writing-style analysis requires broad read access — exactly the kind of permission that creates compliance headaches under GDPR, HIPAA, or sectoral data handling rules.
- Shadow AI integration: Without DLP policies that account for ChatGPT's connected-app feature, users may connect corporate accounts without IT oversight.
- Token persistence: Granted OAuth scopes often persist until explicitly revoked. Users who enable this feature and forget about it leave a standing data pipeline open.
The core issue isn't that OpenAI is doing something unprecedented — it's that each new AI integration expands the perimeter where sensitive data lives outside your control. Security teams need to treat AI assistant integrations with the same governance rigor applied to any third-party SaaS connection.
Shield53 Recommendations
- Audit AI integrations now. Inventory which users have connected ChatGPT or similar AI tools to corporate accounts. Review OAuth token grants in Google Workspace, Slack admin panels, and Notion admin settings.
- Update acceptable use policies. Explicitly address whether employees may connect corporate accounts to consumer AI services. Don't assume existing policies cover this — they often predate these features.
- Implement DLP controls for connected apps. Ensure your data loss prevention tooling can detect and block OAuth grants to AI platforms. Google Workspace and Slack both offer admin controls to restrict which apps users can authorize.
- Plan for revocation workflows. Build a repeatable process for revoking AI-service OAuth tokens when employees offboard or when a feature like this expands scope unexpectedly.
- Monitor OpenAI's enterprise posture. If your organization uses ChatGPT Enterprise or Team, engage OpenAI's support to understand whether Writing Style will be available in managed tenants and what controls administrators will have.
- Educate users. Many employees will see this as a productivity boost and enable it without understanding the data exposure. Brief training on the risks of connecting work accounts to consumer AI tools is worth the effort.
Broader Implications
This feature is part of a larger pattern: AI assistants are evolving from tools that process what you give them into platforms that reach into where your data already lives. That's a powerful convenience, but it shifts more of the security burden onto the permissions model and governance practices surrounding connected accounts. As these integrations become standard, the organizations that treat them as a managed surface — rather than an afterthought — will be the ones that avoid painful data exposure incidents.
For now, this is a limited test. But the trajectory is clear. Security teams should be preparing for a near future where AI assistants are persistently connected to the collaboration tools that hold an organization's most sensitive communications.
