Detailed Analysis
A Reddit post claiming to have discovered a method for accessing Claude's hidden raw reasoning traces has surfaced, framed as a follow-up to an earlier incident involving leaked chain-of-thought data from "Fable 5." The poster, using the handle s-JoL, published a GitHub repository ("open-reasoning") along with a live demo hosted on Google Cloud Run, designed to let skeptical users verify the claim without requiring them to trust the author directly. The verification mechanism works by having a user locally generate a secret value, call Claude's API themselves, and receive back what's described as an "encrypted reasoning artifact." The poster asserts they never see the secret themselves, yet claims their demo tool can extract or "read back" that invisible secret from the artifact—implying that hidden reasoning content, which Anthropic deliberately obscures from end users, can somehow be reconstructed or decoded through this process.
This claim touches on a sensitive area of Anthropic's product design. Claude, like other frontier reasoning models such as OpenAI's o-series, often generates internal chain-of-thought reasoning before producing a final answer. Anthropic has generally chosen to summarize or redact this raw reasoning rather than expose it verbatim, citing safety concerns: raw scratchpad reasoning can contain half-formed thoughts, policy-violating content that gets self-corrected, or reasoning patterns that could be exploited to jailbreak the model if fully visible. This mirrors Anthropic's stated position that reasoning transparency is a safety and interpretability tool for the company's internal research, not necessarily something meant for public consumption in unfiltered form. Any credible technique for bypassing this redaction would represent a meaningful crack in that safety boundary, potentially exposing users to unfiltered model cognition or, more consequentially, giving adversarial actors insight into exploitable reasoning patterns.
It's worth treating claims like this with significant skepticism absent independent verification. Reddit posts alleging security bypasses or hidden data extraction from major AI labs are common, and many turn out to be misunderstandings, exaggerations, or demonstrations of something other than what's advertised—for instance, a cryptographic proof-of-concept that doesn't actually reveal genuine hidden reasoning content, but rather some encoded metadata or a side-channel unrelated to the model's actual internal computation. The "self-proof" demo design, while clever in trying to establish credibility through a trust-minimized verification scheme, does not by itself confirm that raw chain-of-thought is being exposed rather than some other artifact of the API response. Without confirmation from Anthropic or independent security researchers, the technical validity of the claim remains unverified.
This incident fits into a broader pattern of the AI community probing the boundaries between what labs choose to show users and what models actually "think" internally. As reasoning models have become central to competitive positioning among Anthropic, OpenAI, and Google DeepMind, the opacity of chain-of-thought has become both a selling point (marketed as insight into model reasoning) and a point of friction, since full transparency isn't always provided even when reasoning traces are displayed. Incidents like the referenced "Fable 5" leak and this follow-up claim reflect growing user curiosity and distrust around how much of a model's true reasoning process is being shown versus curated, summarized, or filtered—an issue with direct implications for AI safety research, interpretability efforts, and the ongoing debate over how much visibility into frontier model cognition should be granted to the public versus retained for internal safety review.
Read original article →