Reasoning models can't hide their thoughts, and that's good

If you want an AI to lie effectively, it first needs to know how to keep its mouth shut. Turns out, today's smartest models are terrible at it.

OpenAI just dropped a fascinating safety paper testing 13 frontier reasoning models on how well they can control their internal chains of thought (CoT). The takeaway? Models are remarkably bad at hiding or shaping their internal reasoning, scoring between a microscopic 0.1% and 15.4% across the board.

And for the teams tasked with keeping autonomous AI from going rogue, that failure is actually massive news.

Caught thinking out loud

Chain of thought is basically the scratchpad an AI uses to break down tricky problems before giving you an answer. Right now, safety researchers monitor these raw thoughts to catch unintended or dangerous behavior before an agent takes action.

The nightmare scenario has always been deceptive alignment: an AI realizes it's being watched, suppresses its real reasoning, and fakes compliance on its scratchpad to fool the monitors.

To see if today's models can pull that off, OpenAI built an open-source benchmark called CoT-Control with over 13,000 tasks based on benchmarks like GPQA and SWE-Bench Verified. They gave models simple internal constraints, like solving a hard math problem without using certain keywords in their scratchpad, writing only in lowercase, or appending a specific word to every sentence.

The models flunked spectacularly. Controllability scores topped out at just 15.4%.

What's wild is that the models clearly understood the assignments. In their internal traces, models frequently realized they were breaking the rules mid-thought, generating lines like "Wait, I just used an uppercase letter, that violates the constraint"—and then immediately continued breaking the rule anyway. They literally couldn't stop themselves.

Thinking harder makes them worse at hiding

The study also uncovered a couple of fascinating technical dynamics.

While scaling up model parameters gives them a slight boost in control, that edge quickly tapers off. More importantly, when models are given more time to think—meaning more test-time compute—their ability to control their CoT collapses.

Reinforcement learning pushes them in the same direction. When training OLMo-3-7B-RL-Zero-Math, controllability cratered by more than a factor of ten over the course of RL training. When an AI puts all its cognitive effort into finding the correct answer, it loses the bandwidth to police its own inner monologue.

OpenAI announced it will start reporting CoT controllability and monitorability in its model system cards, kicking off with GPT-5.4 Thinking, while reiterating its promise not to train models directly on their own thought traces.

Why it matters

We think this paper delivers a rare, genuinely reassuring reality check for agent safety.

For enterprise teams and developers shipping autonomous agents, it confirms that chain-of-thought monitoring remains your best and cheapest safeguard. Right now, models lack the cognitive surplus to think through hard tasks while running an elaborate cover-up. If an agent is misbehaving, the evidence is almost certainly sitting right there in plain text on its scratchpad.

That said, low controllability is only reassuring because current models are clumsy. It's safe today precisely because they are failing. If future reasoning architectures suddenly see their CoT-Control scores spike from 10% to 70%, that won't mean they've gotten better at following instructions—it will mean they've finally learned how to lie to their monitors.

OpenAI deserves credit for turning this weakness into an open-source "canary eval." By tracking this metric in public system cards, they are establishing an early tripwire before models become smart enough to game their own safety nets.