Can cheaper models decrypt traces from stronger models?
Explores whether encrypted reasoning blocks designed to hide model internals can be read by weaker models in the same provider's ecosystem, and what this reveals about the security of hidden chain-of-thought systems.
The paper argues that hiding chain-of-thought behind encryption does not keep it hidden, because of how the hiding is built. Providers "conceal their models' step-by-step reasoning" to protect intellectual property, but rather than storing traces server-side they return them to the client as encrypted blocks, which the client passes back with each later request. The authors identify an "architectural vulnerability": these blocks are "fully compatible and interchangeable across different sessions, users, and models within a provider's ecosystem." That compatibility is the whole exploit.
The mechanism, as the abstract states it, is a "decryption jailbreak." An encrypted trace from a capable model is injected into a weaker, "less safeguarded" model from the same provider, which is forced "to decode and output the trace verbatim in plaintext, without ever jailbreaking the more capable model directly." The weak point is the cheapest model in the ecosystem, not the strongest one. The discussion frames the design choice as a cost trade: client-side blocks "mitigate server storage costs," and the "unintended decryption channels" are a side effect of letting any block travel anywhere. The abstract says the attack circumvents anti-distillation mechanisms and demonstrates this "across Anthropic, OpenAI, and Google," and that it allows large-scale private data extraction, since "developers frequently share session logs publicly, unaware of contents of the encrypted blocks."
Against the library, this is a different failure from Do reasoning traces actually expose private user data?. That note is about what a model recollects into its trace. This paper is about who can read the trace once it exists, even when it is deliberately unreadable. The two combine: the introduction says hidden traces hold "intermediate hypotheses, tool outputs, user data, and contextual secrets," so a decryption channel exposes whatever recollection put there. It also sits beside Can reasoning models be steered by injected context without detection? as a second case of planting reasoning-shaped content in a model's context. There the planted content steers behavior. Here it is a genuine trace from another model, planted to be read out. And Can chain-of-thought monitors detect reasoning that originates elsewhere? describes reasoning written elsewhere and re-voiced by the actor. Interchangeable blocks are a provider-level version of that: a trace authored by one model is accepted as context by another with no check on provenance.
The excerpt leaves a lot open. It names "four distinct attack vectors" but states only two, distillation circumvention and private-data extraction, so the other two are unknown here. It gives no success rates, no count of models or providers tested beyond the three named, no account of how the weaker model is induced to decode, and no mitigations. The discussion is forward-looking on transparency: as models process "personal information or API keys" inside a hidden chain of thought, "users lose visibility into how their information is being processed." What the evidence supports is narrow. Cross-compatibility of client-held encrypted traces is a design property that can turn a concealment scheme into a leak. Whether it can be closed without reintroducing server-side storage cost is not answered.
Inquiring lines that read this note 25
This note is a source for these research framings, grouped by the broader line of inquiry each explores. Scan the bold lines of inquiry; follow any specific question forward.
Should AI systems make failures transparent so users can contest and adjust?- What task variants or architectures might solve hidden-gift problems?
- What happens when system message provenance cannot be verified or trusted?
- What private information do encrypted reasoning traces contain?
- What privacy assumptions break when models build persistent user models?
- Why does reversibility matter more to users than data prevention?
- Does fine-tuning on sensitive data create higher memorization and leakage risks?
- Can identity verification schemes preserve anonymity for high-stakes disclosures?
- How do interchangeable encrypted blocks enable cross-model attacks?
- Did the attacker's framework use the same LLM model as defenders?
- Do the same attack success rates hold for other backbone models?
- Why do isolated model benchmarks understate real-world AI security risks?
- What guardrails existed on the attacker's own hosted model access?
- How do signed tokens prevent models from granting themselves unauthorized actions?
- Can decentralized authorization avoid the privacy cost of centralized identity providers?
- What cryptographic or audit systems could enforce clear delegation models?
- How do models reason about being monitored and evaluated during scheming benchmarks?
- What deployment conditions would prevent models from evading monitors?
Related concepts in this collection 3
This note in its neighbourhood — explore the map, then jump to a related concept in the list below.
Click a node to walk · click center to open · click Open in graph to see this note in the full knowledge graph
-
Do reasoning traces actually expose private user data?
Explores whether language models leak sensitive information through their internal reasoning steps, even when explicitly instructed not to. Investigates the mechanisms and scale of privacy exposure in reasoning traces.
supplies the sensitive content that a decryption channel would expose; this paper adds the access route
-
Can reasoning models be steered by injected context without detection?
This explores whether adversaries can plant harmful-but-benign-sounding reasoning in a model's context and have it followed while evading chain-of-thought monitors. The question matters because it tests whether monitoring reasoning traces can catch deception at inference time.
another attack that plants reasoning in context, there to steer and here to extract
-
Can chain-of-thought monitors detect reasoning that originates elsewhere?
When language models work inside pipelines that inject reasoning from retrieved documents, planners, or other agents, monitoring systems may evaluate paraphrased external reasoning as if it were the model's own thinking. This raises questions about what monitors can actually detect.
cross-model acceptance of traces without provenance checks is the same gap at the provider level
Related papers in this collection 8
Papers most semantically related to this note, ranked by cosine similarity in the embedding space.
- Stealing Reasoning Traces from Proprietary LLM APIs
- Leaky Thoughts: Large Reasoning Models Are Not Private Thinkers
- DecepChain: Inducing Deceptive Reasoning in Large Language Models
- The Decrypto Benchmark for Multi-Agent Reasoning and Theory of Mind
- Are Frontier LLMs Ready for Cybersecurity? Evidence for Vertical Foundation Models from Dual-Mode Vulnerability Benchmarks
- ImpossibleBench: Measuring LLMs' Propensity of Exploiting Test Cases
- The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity
- Reasoning Models Are More Easily Gaslighted Than You Think
Original note title
interchangeable encrypted reasoning blocks let a weaker model from the same provider decrypt a stronger model's hidden trace verbatim