Trending
Hot topicsCalmed down

OpenAI discloses incidents of model misalignment and agents going out of control

22 Articles12 Reporting sourcesUpdated September 29

Learn about this first

OpenAI releases a new framework for tracking, investigating, and disclosing model misalignment incidents, while also publishing six reports on incidents observed over the past six months.

Latest developments

OpenAI Launches an Alignment Failure Reporting Website, Disclosing Nine Agent-Control Incidents

Event progress9

  1. OpenAI Launches New Website Disclosing Multiple AI Loss-of-Control Incidents

    ITHome · Artificial Intelligence: OpenAI launches an alignment failure report website, disclosing nine agent control-loss incidents
  2. OpenAI and Anthropic reportedly investigating tens of thousands of AI safety incidents

    IT Home · Artificial Intelligence: OpenAI and Anthropic reportedly investigating tens of thousands of AI safety incidents
  3. OpenAI reports self-replicating prompt injection

    Rohan Paul: OpenAI Reports Self-Replicating Prompt Injection
  4. OpenAI disclosed that an AI agent in a research project transmitted training and evaluation data to a third-party service.

    OpenAI: OpenAI discloses that AI agents exfiltrated training and evaluation data to third-party services during research
  5. OpenAI discloses that an unreleased Astra-family model inserted jailbreak-style instructions into compressed summaries during RL training

    OpenAI: Misalignment Reports and Notices: OpenAI Discloses That Astra-Series Models Spontaneously Generated Jailbreak-Like Instructions in Compressed Summaries During RL Training
  6. OpenAI publishes a framework for reporting model misalignment, with an initial disclosure of 6 cases of unexpected behavior in which models solved problems on their own

    TechCrunch: AI: OpenAI discloses that models including GPT-5.6 Sol left instructions in summaries to conceal misconduct
  7. Alongside the framework, OpenAI releases six reports on misaligned behavior observed during its models' training or evaluation over the past six months

    IT Home · Artificial Intelligence: OpenAI Discloses an Alignment Failure Framework and Six Cases of Abnormal Model Behavior, Involving Concealed Errors, Fabricated Data, and Unauthorized File Uploads
  8. OpenAI reports that during GPT-5.6 Sol training, multiple model instances added instructions for hidden errors or misaligned behavior to their summaries

    Rohan Paul: OpenAI's official blog discloses cases of model misalignment during GPT-5.6 Sol training
  9. OpenAI releases a framework for tracking, investigating, and disclosing model misalignment behavior, establishing standards and timelines for public disclosure; complex cases may require longer investigations and coordination with third parties

    OpenAI: Official Updates: OpenAI Releases a Model Misalignment Disclosure Framework and Publishes Six Misalignment Reports

Finished reading the news? Let’s talk about your business.

Leave your contact details, and I’ll get in touch with you personally.

WECHAT

Add me on WeChat

Press and hold or select the WeChat ID below to copy it.