Short answer: On 30 October 2026 OpenAI emailed users of Daybreak Blue, its trusted-access tier for cyber work, that from 29 November content from Daybreak Blue sessions on personal accounts may be used to improve its models, even with the usual “Improve the model for everyone” setting off. Researchers objected that this could include unreported bugs and proofs of concept. Reports then said OpenAI promised a separate opt-out for Daybreak sessions, without publishing how or when. The practical lesson does not depend on how that ends: decide what may go into an AI tool before you paste it, save a dated copy of the terms you relied on, and keep your own dated record of a finding outside the tool. A TimeBind dates the exact files you keep. It does not stop anyone using what you enter, and it does not show you found a bug first.
Last checked 13 November 2026. General information for security researchers and developers, not legal advice and not advice on any disclosure programme. This page reports what was said publicly in late October and early November 2026. OpenAI’s terms may have changed since, so read its current data-controls page before you rely on anything here.
Table of Contents
What was reported
According to a report by newruntime.com that cites CellCog, OpenAI emailed Daybreak Blue users on 30 October 2026. From 29 November, the email said, OpenAI may use content from Daybreak Blue sessions on personal accounts to improve its models, even if the account setting “Improve the model for everyone” is off. The report says OpenAI’s data-controls help page gained a matching section on “Trusted Access on personal accounts,” that the rule applies during Daybreak sessions while the usual opt-out still applies to other activity, and that business and enterprise customers are not affected. Daybreak is OpenAI’s Trusted Access for Cyber programme; the Blue tier offers reduced refusals for work such as vulnerability triage, secure code review and malware analysis, as the same report describes it.
Security researcher Sean Heelan wrote on X that training on Daybreak Blue sessions with no way to opt out was “cartoonishly evil” and would likely end Daybreak use by security researchers. Others on X said their sessions mostly contained other people’s unpatched bugs that they were not allowed to share.
After the backlash, RuntimeWire reported (updated 31 October) that a later notice promised a separate opt-out for Daybreak sessions, and newruntime.com and AI.info reported the same promise. As those reports describe it, the interface, default state and timing had not been published, with more detail expected near 29 November. We did not find OpenAI’s own statement of the final rule, so check its current data-controls page.
Why this matters even if you never use Daybreak
The same question applies to any AI tool: which account, which programme tier and which setting decides whether what you type can be used for training? The newruntime report makes the point that a global opt-out is not necessarily the operative boundary for a specialised access programme. For researchers, the stakes are specific: unreported bugs, reproducers and client targets are often under embargo or under contract.
Whatever a vendor decides, you can protect your own position by keeping a dated record of your finding and of the terms you relied on, outside the tool.
Why people are searching for this
Measured on 13 November 2026. On X, Sean Heelan’s post about the plan (30 October) had 208,293 impressions and 1,494 likes, and a reply from another security researcher confirming it (30 October) had 163,089 impressions and 847 likes. Other posts on the same story had 66,913, 18,066, 16,461 and 16,328 impressions (30 October). Google autocomplete in the United States returned “openai daybreak blue,” “daybreak blue openai access” and “openai trusted access cyber” on 13 November. Counts keep moving.
Checklist for security researchers and developers
- Decide what may enter an AI tool before you start. Write a short list: unpatched vulnerabilities, proofs of concept, client identifiers and source fragments may need an approved, contract-safe route, or no AI tool at all.
- Check the account type, programme tier and data-use rule. The reports describe a rule that applies to one tier on personal accounts. Read the current terms for the exact combination you use.
- Save a dated copy of the terms you relied on. A PDF or screenshot of the help page and settings, saved on the day you started a piece of work, shows what the rules said then.
- Keep your own record of each finding outside the tool. Notes, the proof of concept, tested versions and the date you found it, saved before you share it anywhere. Date the archive.
- Keep the disclosure timeline. The date you reported it, any acknowledgement, the patch and the public date, as dated emails or tickets.
- Log AI-tool sessions that touched a finding. Tool, account, date and a summary of what was entered, without copying the sensitive content into the log.
- Follow your contract and programme rules. Your NDA, client agreement or bug-bounty terms may limit where findings can go. If you are unsure, ask in writing and keep the reply. See also what to keep before you install an AI tool and GitHub, Copilot and Cursor training opt-outs.
What a TimeBind does and does not show
A TimeBind is a dated, checkable record that an exact file existed no later than a recorded time and still matches. The SHA-256 fingerprint is made in your browser and the file is never uploaded. The record is anchored to Bitcoin (a Merkle root in an OP_RETURN). Your first Pooled TimeBind is free; after that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. See how it works.
For a finding, it can date the report, the proof-of-concept archive and a saved copy of the terms you relied on. It does not stop anyone using what you enter in a tool, does not show that you found a bug first, and does not decide credit, bounty or disclosure rights. It helps you show which files you held by a given time and that they have not been altered since.
Limits
This page relies on newruntime.com (reporting an email OpenAI sent on 30 October 2026, read 13 November), RuntimeWire (30 and 31 October), AI.info and search summaries (seen 13 November), and the X and autocomplete counts above. We did not see the email itself or OpenAI’s current help page. The terms may have changed. This is not legal advice.
Tools that can help
Use the bulk SHA-256 manifest to fingerprint a folder holding a report, a proof of concept and saved terms, then keep a dated copy of the manifest.
Frequently Asked Questions
-
What did OpenAI tell Daybreak Blue users?
According to reports of an email sent on 30 October 2026, that from 29 November content from Daybreak Blue sessions on personal accounts may be used to improve its models, even with the usual setting off.
-
Did OpenAI change the plan?
Reports from 31 October say a later notice promised a separate opt-out for Daybreak sessions, but how and when were not published. Check OpenAI’s current data-controls page for the final rule.
-
Are business accounts affected?
The newruntime report says business and enterprise customers are not affected by the rule it describes. Check your own terms.
-
Can a TimeBind stop an AI company using what I enter?
No. It does not control how any tool uses your input. It dates the files you hold.
-
Can a TimeBind show I found a bug first?
No. It shows that exact files existed no later than a recorded time and still match. It does not show who found a flaw first.
Sources checked
- newruntime.com, “OpenAI will train on personal-account Daybreak Blue sessions from November 29 even with the opt-out off,” https://newruntime.com/posts/openai-will-train-on-personal-account-daybreak-blue-0571d8/ (read 13 November 2026)
- RuntimeWire, Ryan Merket, “OpenAI walks back Daybreak Blue training plan for personal accounts,” 30 October 2026, updated 31 October, https://runtimewire.com/article/openai-daybreak-blue-training-opt-out-api-access (headline and summary read 13 November 2026)
- AI.info and newruntime.com reports of the promised opt-out (seen in search results 13 November 2026)
- X posts by Sean Heelan and others (30 October 2026) and Google autocomplete (US), checked 13 November 2026
Date the report and saved terms you keep yourself
Your first Pooled TimeBind is free, with no credit card required. After that, TimeBinds start from US$5, and an ASAP TimeBind is US$15. Click here to start, or see plans and pricing.
