Discovering unknown AI misalignments in real-world usage

Reasoning models can find and understand unknown misaligned behaviors from how users respond.

Discovering unknown AI misalignments in real-world usage ← Back to OpenAI Alignment Blog Discovering unknown AI misalignments in real-world usage Jan 2026 · Hannah Sheahan Reasoning models can find and understand unknown misaligned behaviors from how users respond. As AI systems become more capable and reach a broader audience, the range of real-world interactions—and the failures that accompany them—will continue to expand. Despite extensive testing, there is a limit to what we can learn in a lab before deployment. Inevitably, “unknown unknowns” arise only once models are exposed to the full

related reading

• Current AIs seem pretty misaligned to me — LessWronglesswrong.com
• Current AIs seem pretty misaligned to meblog.redwoodresearch.org
• How confessions can keep language models honest | OpenAIopenai.com
• Teaching Claude why \ Anthropicanthropic.com
• Natural-emergent-misalignment-from-reward-hacking-paper.pdfassets.anthropic.com
• [2502.17424] Emergent Misalignment: Narrow finetuning can produce broadly misaligned LLMsarxiv.org
• Alignment will happen by default. What’s next? — LessWronglesswrong.com
• Agentic Misalignment in Summer 2026alignment.anthropic.com
• Model Organisms of Misalignment: The Case for a New Pillar of Alignment Research — AI Alignment Forumalignmentforum.org
• Alignment faking in large language modelsarxiv.org
• How well do models follow their constitutions? — LessWronglesswrong.com
• Discovering Language Model Behaviors with Model-Written Evaluations — LessWronglesswrong.com