OpenAI's own data-controls help page now carries this sentence, which we verified directly today: during "Daybreak" sessions on a personal account - its gated program for cybersecurity professionals - the company "may use content from your interactions to improve our models even when 'Improve the model for everyone' is off." Separately, emails sent to program members and reported by The Decoder say the change takes effect November 29; the public help page itself names no date.
This affects a narrow group of users, and it is a stated possibility ("may use"), not a declaration that every session will be used. But the principle behind it should get every business owner's attention, because it demolishes a comfortable assumption: that a settings toggle is a guarantee.
What actually changed
Daybreak is OpenAI's "Trusted Access for Cyber" program - application-only access for people doing authorized defensive security work. The Blue tier, the starting point for most teams, covers secure code review, vulnerability triage, malware analysis and incident response. Individual access requires a paid plan, advanced account security, and a hardware security key.
Until now, those sessions were excluded from training whenever the user had disabled training. The updated help page carves out an exception for Daybreak sessions on personal accounts. OpenAI's stated reason: these sensitive interactions are exactly what helps it decide what to block or allow, strengthening defenses against malicious use of AI.
After swift backlash, OpenAI sent a follow-up email promising a separate opt-out mechanism, per The Decoder's reporting. Interactions flagged during safety reviews can still be used for safety purposes regardless.
Two things are worth noting for accuracy. Business, Enterprise, Edu and healthcare workspaces - and API usage - remain excluded from training by default. And the November 29 date comes from user-reported emails; the public help page itself names no date.
Why this matters to a business that has never heard of Daybreak
Most Indian small businesses will never touch a cybersecurity program like this. The lesson is not about Daybreak - it is about how AI providers treat your data as a policy decision they can revisit.
Think about what your team pasted into AI tools this month. A draft client contract. Salary figures for a hiring decision. Source code. A customer's complaint history. If that went into a personal ChatGPT account, its protection depends entirely on a settings page the provider can edit - as this episode just demonstrated. We made the same point when Sweden's government blocked ChatGPT for civil servants last week: the risk was never that AI is bad, it is what employees paste into it without a policy.
What to do about it, practically
- Know which account type your team uses. The protections differ sharply: personal accounts train by default unless toggled off (with exceptions, as we just saw), while Business and Enterprise workspaces are excluded from training. If staff use AI for client work, a workspace plan is the baseline, not a luxury.
- Write a one-page AI data policy. What may never be pasted into any AI tool (client personal data, financials, credentials, unpublished code), which tools are approved, and who answers questions. Sweden's government now runs on exactly this model - one approved tool, clear rules.
- Review AI settings quarterly. Data-use policies are changing every few months across every major provider. Put a recurring check in the calendar; this week's change appeared silently on a help page.
- Treat the toggle as a preference, not a vault. The safest data is data that never leaves your hands. For genuinely sensitive work, anonymize before you paste.
The bigger pattern
Within one week, the biggest AI companies agreed to let websites opt out of AI training through robots.txt, and OpenAI moved a category of user sessions in the opposite direction - from excluded to included, pending an opt-out. The direction of travel is clear: providers will keep redrawing the line between "your data" and "training data," and the defaults will usually favor the provider.
Businesses that rely on a toggle they set once in 2024 are trusting a moving target. A short policy, the right account type, and a quarterly review are cheap insurance.
If you are not sure which AI tools your team is using or what data is flowing into them, that audit is exactly the kind of work we do. VAIR IT Technologies offers a free consultation - we will help you map your AI usage and put a sensible policy around it. Reach us at vairit.com/contact.
