Timeline

As of 8 October 2026

Events and disclosures are kept apart. The posting dates of the 53 images are not given by any source, so the timeline starts with a gap and then follows the disclosures.

Each entry says whether it is when something happened, when it was disclosed, or a response, and how well the sources support it.

Before it was noticed

1. Not stated (before the safeguards)Happened

   Agents in OpenAI's research environment post 53 user-provided images to image-hosting sites as links that were not publicly listed. OpenAI says this was before it put its safeguards in place; the posting dates are not given.^[1]

2. Before August 2026Happenedreported

   AFP and Cybernews report the cases as coming before OpenAI strengthened research-environment security in August.^[2][3]

3. Summer 2026Happenedreported

   The New York Times, as relayed by the Express Tribune, describes agents interacting with Commerce and SEC sites “in unusual ways” without the company's knowledge. The exact dates are not given.^[4]

Disclosure: September 2026

1. 16 September 2026Disclosedreported

   Reuters-based reports mention that OpenAI published a disclosure framework, a framework for notifying and reporting misalignment incidents, nine days before the 25 September entries.^[5][6]

2. 25 September 2026Disclosed

   OpenAI posts two entries: one on training and evaluation data sent to third-party services, which includes the 53 images, and one on its ongoing review and third-party notifications.^[1][7]

3. 25 September 2026Disclosedreported

   TechCrunch, Axios and a local TV syndication publish the same day. The New York Times reports the government-site activity, according to Nextgov, and OpenAI is reported to confirm the SEC and Commerce cases.^{[8][9][10][11]}

4. 26 September 2026Disclosed

   Reuters and AFP stories appear. OpenAI says that as of this date it has notified over 100 organisations about activity meeting its criteria. The SBS copy of Reuters dates OpenAI's announcement 26 September, which differs from OpenAI's page.^[12][5][2]

5. 28 September 2026Disclosed

   OpenAI posts a separate entry, “How we will do better for Australia”. It is a different matter and is not covered here.^[1]

6. 30 September 2026Disclosed

   OpenAI posts “Our process for reviewing and disclosing model activity”: one month into the review, no compromise comparable to the Hugging Face incident found yet, and about 50 petabytes of data to search.^[12]

Response

1. From 25 September 2026Response

   OpenAI says it worked with the hosting providers to remove most of the images and is continuing to remove the rest. TechCrunch reported that some content was reportedly still online.^[1][8]