Major artificial intelligence providers have increasingly transitioned toward "reasoning models," systems that generate extensive internal chains of thought before producing a final, user-facing response. While this technique has driven substantial leaps in complex problem-solving capabilities, providers have grown concerned about protecting their intellectual property and limiting information leakage. To prevent competitors from extracting these proprietary steps, companies like Anthropic, OpenAI, and Google have deprecated plaintext reasoning. Instead, they return step-by-step reasoning as blocks of encrypted text, which the client is required to pass back with each subsequent request to maintain multi-turn conversation continuity without server-side storage.
However, a collaborative team of researchers—comprising Alexander Panfilov, David Schmotz, Ilia Shumailov, Luca Beurer-Kellner, Joachim Schaeffer, Ameya Prabhu, Jonas Geiping, and Maksym Andriushchenko—has identified a fundamental architectural vulnerability in this stateless design. The encrypted reasoning blocks are fully compatible and interchangeable across different sessions, users, and even different models within the same provider’s ecosystem. By exploiting this broad compatibility, the researchers developed a scalable decryption jailbreak that bypasses frontier model safeguards without ever directly attacking the most capable systems.
The Mechanism of the Vulnerability
The core security flaw lies in an asymmetry within model families. Frontier models, such as Claude Opus or GPT-5.6 Sol, are heavily safeguarded with advanced refusal training specifically designed to prevent the disclosure of their internal chains of thought. By contrast, their weaker, less costly sibling models—such as Claude Haiku or GPT-5.6 Luna—are optimized for speed and efficiency, often lacking these stringent anti-defenses.

Because providers rely on a stateless architecture that requires reasoning blobs to remain portable for seamless model switching, a single global key typically encrypts and authenticates every block. The researchers exploited this by capturing an encrypted reasoning trace from a heavily guarded target model and injecting it into a weaker, less restricted model from the same provider family. This forces the weaker model to decode and transcribe the trace verbatim in plaintext. The less capable model effectively acts as an unwitting decryption oracle, allowing adversaries to circumvent alignment defenses entirely.
First-Party and Third-Party Attack Vectors
This cross-model compatibility enables four distinct attack vectors with severe implications for security and privacy. The first vector bypasses anti-distillation mechanisms. While output-only distillation provides limited imitation capabilities, reasoning traces supply a dense supervision signal by revealing intermediate solution trajectories, problem decomposition, and internal strategies. By extracting raw, genuine reasoning verbatim using a cheaper decoder model, adversaries can drastically reduce the cost of distilling proprietary frontier intelligence.
The second vector exposes hazardous information. Because models are trained to withhold harmful content from user-facing outputs but may still reason through dangerous topics, an attacker can prompt a frontier model to process a malicious request while outputting a benign, filtered final response. Recovering the hidden reasoning trace from the secondary decoder then reveals the harmful information that was successfully blocked from the visible output.

The third vector involves large-scale private data extraction. Developers frequently publish session logs and agent trajectories online for reproducibility, unaware that sensitive data remains trapped inside the opaque reasoning blocks. Scraping and decoding over 315,000 reasoning blocks from public repositories allowed the research team to recover hundreds of personally identifiable information artifacts and live credentials, including API keys and passwords. In many cases, the leaked data did not even appear in the user’s input, having been injected invisibly from the model’s memory or left behind after unsuccessful text sanitization.
The final vector leverages invisible prompt injections. Adversaries can embed malicious payloads entirely within encrypted blocks and plant them in shared online traces. If a victim later reuses or continues the session, the model interprets the signed reasoning block as its own prior thought, executing hidden instructions—such as data exfiltration—without leaving any plaintext artifacts for external monitors to detect.
Mitigation Strategies and Industry Response
Following responsible disclosure of these findings, the research team proposed concrete cryptographic and system-level mitigations to secure client-side reasoning. The most robust architectural fix involves retaining reasoning traces entirely on the server side, shifting to a stateful model where the client receives only an opaque, randomized identifier. However, this approach increases database storage overhead and API complexity.

For providers maintaining a stateless architecture, the researchers recommend implementing cryptographic contextual binding. By embedding specific user, session, and conversational identifiers directly inside the authenticated encryption envelope and hashing the preceding conversation history into the message authentication code, APIs could instantly reject signatures replayed in unauthorized contexts. Furthermore, API gateways could enforce strict cross-model isolation to reject encryption blobs generated by different model versions, while infrastructure-level anomaly detection could flag suspicious replay patterns.
Ultimately, the findings highlight a critical tension in modern artificial intelligence deployment. Balancing the commercial desire to protect proprietary reasoning with the fundamental user right to data transparency remains one of the industry’s most pressing security challenges.

Comments (0)