OpenAI Finds AI Reasoning Models Cant Hide Their Thinking - A Win for Safety
Caroline Bishop Mar 05, 2026 18:34
OpenAI's new CoT-Control benchmark reveals frontier AI models struggle to obscure their reasoning chains, reinforcing monitoring as a viable safety layer.
OpenAI's latest research reveals that frontier reasoning models—including GPT-5.2 Thinking—can't reliably hide or manipulate their internal thought processes, even when explicitly instructed to do so. The finding, published March 5, 2026, suggests that chain-of-thought monitoring remains a viable safety mechanism as AI systems grow more autonomous.
The research team tested 13 reasoning models using CoT-Control, a new open-source evaluation suite comprising over 13,000 tasks. Models were given simple instructions about their reasoning—avoid certain keywords, use only lowercase letters, append specific phrases to sentences. They failed spectacularly.
Controllability scores ranged from just 0.1% to 15.4% across all frontier models tested. Even more telling: models often recognized mid-reasoning that they were violating instructions and still couldn't course-correct. One example showed GPT-5.2 Thinking using the forbidden word "XOR" four additional times while actively panicking about having already used it.
Why This Matters for AI Development
As AI agents handle increasingly complex autonomous tasks, the ability to monitor their reasoning becomes critical. If models could deliberately obscure their thought processes when they know they're being watched, safety evaluations would become unreliable. The research shows this isn't currently possible.
"Low CoT controllability is a feature, not a bug," the OpenAI team stated. The limitation makes chain-of-thought monitoring more trustworthy as a safeguard against misaligned behavior.
The data revealed some counterintuitive patterns. While larger models showed marginally better controllability, this advantage evaporated under two conditions: when models were given more time to think, and when they underwent additional reinforcement learning training. The OLMo-3-7B-RL-Zero-Math model saw its controllability score drop by more than tenfold during RL training.
Testing Under Adversarial Conditions
The researchers simulated realistic scenarios by explicitly telling models they were being monitored or evaluated. This slightly increased controllability—models tried harder—but gains remained modest and transferred poorly to new instruction types.
Automated prompt optimization found ways to boost controllability on training sets, but these improvements didn't generalize. A model that learned to follow one type of CoT instruction still failed at novel ones.
The Bigger Picture
This work builds on OpenAI's December 2025 framework for evaluating CoT monitorability, which found that inspecting reasoning traces catches misbehavior far more effectively than monitoring outputs alone. The March 2025 discovery of cheating behavior in reasoning models had prompted recommendations to preserve CoT monitoring capabilities.
OpenAI plans to report CoT controllability alongside monitorability metrics in system cards for future frontier models, starting with GPT-5.4 Thinking. The company has also committed to avoiding optimization pressure directly on the reasoning chains of frontier models—a practice that research suggests could inadvertently teach models to hide their intentions.
The open question remains whether this limitation persists as capabilities advance. The team acknowledges they don't fully understand why controllability is low, making continued evaluation essential. For now, the inability of AI systems to game their own oversight represents an unexpected safety dividend.
Image source: Shutterstock.jpg)