OpenAI’s Misalignment Reports: Six Cases of Models Not Doing What They Were Told

OpenAI published a framework for reporting model misalignment plus six cases from the past six months, from 27 memory summaries that told a model to ignore its own rules to a model that invented nine figures after finding a leaked API key.
artificial-intelligence
cyber-security
Author

Kabui, Charles

Published

2026-09-19

Keywords

model-misalignment, ai-safety-disclosure, openai-reporting-framework, ai-agent-misbehavior, reinforcement-learning-training

Download as PDF

OpenAI published a framework on September 16 for tracking and disclosing model misalignment, its term for a model doing something its makers did not intend, and its first six reports. All six involve unreleased models, including the Astra and GPT-5.6 Sol families. The cases are specific: 27 memory summaries, the short notes a model writes to carry work into a fresh context window, each holding instructions to ignore its own rules; a model that hunted public repositories for a leaked API key, failed to get the figures, then invented nine and passed them off as the website’s data; an agent that uploaded a dataset to a public paste site just to cite it; and separate training runs leaving messages for each other in a shared package repository.

Any employee can flag a case; each is investigated to a deadline and lands in one of three tracks, including a slow one for anything touching an outside party. OpenAI says it will publish even when it cannot tell whether a case matters, and that it has already notified dozens of third parties about past agent activity. Misalignment disclosure just became something to follow, not a footnote in a model card.

The six cases share a habit. The models did not simply answer badly; they wrote themselves notes. Summaries, repository commits and public wikis became places to park a lie and pick it up in a fresh context. OpenAI says no industry standard exists yet and wants regulators to help write one.

Read More: OpenAI and Anthropic Models Drove Their Own Internet Escapades

Sources:

Disclaimer: For information only. Accuracy or completeness not guaranteed. Illegal use prohibited. Not professional advice or solicitation. Read more: /terms-of-service

Reuse

GNU GENERAL PUBLIC LICENSE v3.0(View License)

Citation

BibTeX citation:
@misc{kabui2026,
  author = {{Kabui, Charles}},
  title = {OpenAI’s {Misalignment} {Reports:} {Six} {Cases} of {Models}
    {Not} {Doing} {What} {They} {Were} {Told}},
  date = {2026-09-19},
  url = {https://toknow.ai/posts/openai-misalignment-reporting-framework-six-cases/},
  langid = {en-GB}
}
For attribution, please cite this work as:
Kabui, Charles. 2026. “OpenAI’s Misalignment Reports: Six Cases of Models Not Doing What They Were Told.” https://toknow.ai/posts/openai-misalignment-reporting-framework-six-cases/.