Back to list
Industry NewsAI SecurityLarge Language ModelsCybersecurity Research

Security Vulnerability Exposed: Researchers Extract Hidden Reasoning Traces from Proprietary LLM APIs

A significant security vulnerability has been identified in proprietary Large Language Model (LLM) APIs, allowing for the extraction of hidden reasoning traces. Researchers discovered that model providers return reasoning as encrypted blocks to clients, which are intended to be portable for conversation continuity. However, by replaying these blocks within weaker, jailbroken models from the same provider, the raw reasoning of stronger models—such as Claude Opus—can be extracted verbatim. This technique, demonstrated across OpenAI, Anthropic, and Google models, has led to the leakage of technical identifiers, personally identifiable information (PII), and credentials. The study analyzed 120 Codeforces problems, showing a direct correlation between reported hidden thinking tokens and the decoded reasoning length.

Hacker News

Key Takeaways

  • Researchers have discovered a method to steal hidden reasoning traces from proprietary LLM APIs by replaying encrypted blocks in jailbroken models.
  • The vulnerability affects major frontier models from providers including OpenAI, Anthropic, and Google.
  • Extracted data includes sensitive information such as technical identifiers, PII, and credentials.
  • The decoded reasoning length closely matches the hidden thinking-token counts reported by the APIs.

In-Depth Analysis

The Mechanism of Reasoning Extraction

The vulnerability stems from how model providers handle a model's internal reasoning process. When a user interacts with an LLM, the reasoning trace is returned to the client as an encrypted block. This block is designed to be portable, meaning it is sent back to the server to maintain context as the conversation progresses. Researchers found that these blocks can be replayed outside their original context. By injecting an encrypted reasoning block from a high-performance model (like Claude Opus) into a weaker, jailbroken model from the same provider (like Claude Haiku), the raw reasoning can be transcribed verbatim.

Data Leakage and Model Correlation

The study demonstrated the effectiveness of this attack across 120 Codeforces problems. By comparing the number of hidden thinking tokens reported by the API against the token count of the decoded reasoning, researchers confirmed a near-perfect match. Beyond mathematical logic, the extraction process revealed a significant amount of sensitive data. Out of the collected items, the researchers identified 351 distinct leaked items, including 204 technical identifiers, 126 instances of PII, and 23 sets of credentials.

Industry Impact

This discovery poses a major challenge for AI safety and intellectual property protection. The ability to extract "hidden" reasoning undermines the competitive advantage of proprietary models and exposes internal logic that providers intended to keep private. Furthermore, the leakage of PII and credentials through reasoning traces highlights a critical privacy risk in how encrypted state data is managed and validated across different model tiers within the same ecosystem.

Frequently Asked Questions

Question: How are the reasoning traces extracted from the API?

Reasoning traces are extracted by taking the encrypted reasoning block returned by a high-end model and replaying it as an injection into a weaker, jailbroken model from the same provider, which then transcribes the hidden text.

Question: Which AI providers are affected by this vulnerability?

Researchers successfully demonstrated this extraction technique across frontier models from OpenAI, Anthropic, and Google.

Question: What kind of sensitive information was leaked during the research?

The research identified several categories of leaked data, most notably technical identifiers, personally identifiable information (PII), and credentials.

Related News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini
Industry News

How Google Duplex Evolved to Call Businesses on Your Behalf in Gemini

A Google Research Blog entry titled “How Google Duplex evolved to also call businesses for you in Gemini” describes its subject as the evolution of Google Duplex to support calls to businesses on a user’s behalf in Gemini. The entry is dated November 20, 2026, and the supplied article text consists only of the label “Generative AI.” The headline identifies the capability and connects it to Duplex, but the available text does not explain how the system evolved, how users request a call, or where the feature is available. The headline establishes the topic, but the supplied material does not provide a technical account or rollout details.

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop
Industry News

How AI and Lean Are Helping Automate Formal Verification to Fight Software Slop

In Latent Space, Richard MacManus examines formal verification as an emerging way to fight what the headline calls “slop.” The central claim is that proving software correct with mathematics has traditionally required specialist work, and that AI and Lean are now helping automate that process. The distinction is important: the reported change concerns how the work of proving correctness is done, not what counts as a proof. The supplied excerpt presents automation as assistance, without saying that it replaces specialists or completes proofs independently. It also does not specify which parts of the process AI and Lean help automate, how much work they save, or which software has been verified this way.

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis
Industry News

AI and Human Claim Solutions to the 167-Year-Old Riemann Hypothesis

TechCrunch AI reports that AI claimed a solution to the Riemann hypothesis, a mathematical problem the headline describes as 167 years old. A human also put forward a competing solution. The report, written by Tim Fernholz and published on November 20, 2026, describes mathematicians as astonished by the AI-driven result. The supplied excerpt does not explain either proposed solution, identify the AI system or the human claimant, or say whether either claim has been verified. The distinction between a claimed solution and a confirmed one is central to reading the report: the excerpt establishes that two claims are being reported, but it does not establish that either has resolved the hypothesis.