A massive security research effort has exposed a hidden vulnerability running through the world's most prominent AI chatbots. According to a new report, researchers found that a single shared encryption key protects the private reasoning tokens of every major AI provider — and they successfully used it to unlock hundreds of thousands of secret thinking blocks.

The findings, published Wednesday, reveal a systemic weakness in how AI companies protect their models' internal thought processes. The researchers exploited the flaw to decode 315,320 hidden thinking blocks from public logs, recovering sensitive material that included passwords and live API keys.

The Flaw: One Key to Unlock Them All

AI models often generate a hidden chain of reasoning before producing a final answer. These reasoning tokens, sometimes called "inner thoughts" or "thinking blocks," are supposed to remain private. In theory, even if a user sees the final output, they should never see the internal deliberation that led to it.

The researchers discovered that this privacy barrier was fundamentally broken. Instead of each AI provider using unique encryption keys — or each model using its own key — the entire ecosystem relied on the same global key. That meant once the key was identified, it could decrypt reasoning tokens across all major AI providers without requiring any special access or per-target hacking.

How the Exploit Worked

The exploit did not rely on a traditional vulnerability like a buffer overflow or a SQL injection. It was a cryptography and key-management failure. By analyzing public logs and samples of encrypted reasoning tokens, the researchers identified the shared key and then used it to unlock hidden data at scale.

In total, the team decoded 315,320 hidden thinking blocks from public sources. This was not a single model's private cache; it spanned the major AI providers that dominate the industry.

Recovered Secrets: Passwords and Live API Keys

The impact of the exploit goes far beyond reading a chatbot's private thought process. Among the decrypted reasoning blocks, the researchers found real-world secrets that users or other software had fed into the models. That included passwords and live API keys — credentials that, in the wrong hands, could be used to access services, spend credits, or impersonate users.

Because reasoning tokens are sometimes logged for training, debugging, or safety analysis, these hidden evaluations can contain sensitive data. The single global key made all of that data readable by anyone who knew the key.

The discovery highlights a dangerous reality: the same feature that lets AI models think step-by-step can become a leaky archive of user secrets when upstream encryption is mishandled.

What This Means for AI Security

This flaw is a wake-up call for AI providers that rely on centralized, shared infrastructure for protecting model internals. A single key reused across an entire industry violates the basic security principle of unique keys per application and environment. The researchers' ability to decode more than 300,000 reasoning blocks demonstrates that this is not a theoretical risk.

For enterprises and developers, the implications are clear:

  • Do not assume AI reasoning is private. Even when providers promise confidentiality, the underlying key management may be flawed.
  • Treat any data sent to an AI model as potentially exposed. That includes passwords, API keys, and other secrets that might appear in prompts or context.
  • Demand clear encryption and key-management policies from AI providers before feeding them sensitive information.
  • Rotate credentials regularly. If a reasoning-token leak occurs, affected passwords and API keys need to be invalidated quickly.

Who Is Responsible?

The report does not single out one vendor, but it describes a systemic issue affecting "every major AI provider." That means the problem is industry-wide. It also suggests that the typical approach of protecting model outputs and internal states has been dangerously homogenous, leaving a single point of failure for the entire AI sector.

Security researchers have long warned that cryptographic key reuse is a serious design flaw. This case turns that warning into a concrete example with a shocking scale: more than 300,000 leaked reasoning blocks and a cache of credentials that could fuel further attacks.

Key Takeaways

The exposure of hidden AI thinking is not just an academic curiosity. It is a security incident with real consequences. The researchers found that a single global encryption key was enough to break the confidentiality of reasoning tokens across major AI providers, leading to the recovery of passwords and live API keys.

  • A single shared encryption key protected reasoning tokens across major AI providers.
  • Researchers decoded 315,320 hidden thinking blocks from public logs.
  • The decrypted data included passwords and live API keys.
  • The exploit underscores the need for unique, per-application encryption keys and better key management in AI systems.

Until AI providers tighten their cryptographic practices, any data passed to a large language model could be one broken key away from exposure.