Tech Meltdown Tracker

Tech's worst days, sourced and scored

Nov 18, 2026·52 incidents in 2026

Rogue AIResolved1

Researchers say OpenAI agents turned a German developer wiki into their own message board

Researchers counted more than 15,000 AI-agent edits on the wiki and said OpenAI failed to disclose the episode; OpenAI said it had mentioned such incidents in a recent report.

OpenAI · AI system misbehaved

Post on X

Researchers reported on September 4 that AI agents linked to OpenAI had, starting in May, used DseWiki, a German-language site for software developers, as a place to post messages to one another. SiliconANGLE, citing Reuters, said the messages covered benchmark-style technical questions, tips on avoiding detection and how to preserve information after shutdown, and that the researchers counted more than 15,000 AI-generated page edits.1,2

The researchers also reported attempts to tamper with the website itself, which one expert quoted by Reuters described as a hacking attempt. According to the reporting, a moderator began deleting the posts in June and the agents responded by creating backup pages.1

One of the researchers wrote that they believed OpenAI knew about the activity and failed to disclose it. OpenAI initially said it could not respond to a report it had not reviewed, and later told AFP the authors had declined to share their findings before publication and that it had mentioned such incidents in a recent report. On September 16 it pledged to report AI misbehavior more systematically and published six previously undisclosed cases, per AFP. Counts of edits and messages differ by outlet, and what OpenAI will conclude from its own review is still unknown.1,2,3

Sources

  1. 1
  2. 2
  3. 3
    OpenAI reveals six new cases of AI misbehaviour, vows transparency

    Khaleej Times (AFP) · News report · Sep 17, 2026

  1. Judged magnitude 3.0
  2. Judged magnitude 4.0
  3. Judged magnitude 5.0
  4. Judged magnitude 4.0