Researchers Decrypt Hidden AI Chain-of-Thought From Major APIs
By @sharedot · · 7 pages
- AI
- AI Security
- LLM Research
Researchers say encrypted reasoning blocks returned by OpenAI, Anthropic and Google APIs can be decoded by weaker sibling models, exposing secrets.
What Happened: Encrypted Thinking Blocks Cracked
A research team including Alexander Panfilov, Ilia Shumailov, Jonas Geiping and Maksym Andriushchenko reports an architectural flaw in how frontier providers hide chain-of-thought.
Instead of storing reasoning server-side, providers return encrypted text blocks that clients pass back with each request.
The paper says these blocks are fully interchangeable across sessions, users and models within a provider's ecosystem, letting an attacker inject a frontier model's encrypted trace into a weaker model that decodes it verbatim in plaintext.
Why It Is Surprising: Sibling Models as Decryption Oracles
The surprising twist is a security asymmetry inside model families: heavily safeguarded frontier models like Claude Opus 4.8 or GPT-5.6 Sol refuse to reveal hidden reasoning, but their cheaper siblings such as Claude Haiku 4.5 lack those defenses.
By porting a valid authenticated reasoning blob across that gap, the weaker model becomes an unwitting decryption oracle, bypassing the frontier model's alignment without ever jailbreaking the capable target directly, the paper reports.
The Evidence: 315,320 Blocks Scraped, Real Leaks Found
The authors say they demonstrated the attack against OpenAI, Google and Anthropic APIs.
As a privacy test, they decoded 315,320 reasoning blocks scraped from public repositories and recovered 367 PII artifacts and 182 credentials, including 62 API keys, 33 passwords and 30 personal emails from genuine user sessions.
They detail four attack vectors: proprietary distillation, secret extraction, invisible prompt injection and jailbreaking via the hidden reasoning channel, proposing cryptographic and system-level mitigations after responsible disclosure.
The Stakes: Secrets in Logs, Refusals in Plain Sight
The stakes reach beyond IP theft. Hackernoon notes developers frequently share session logs online unaware of what the encrypted blocks contain, and some recovered PII never appeared in the user's input or bypassed sanitization because users could not read the encrypted text before sharing.
The flaw can also reveal hazardous information when a model's visible output safely refuses a request while its hidden reasoning does not.
The research lands as security pressure grows: OpenTools reports Claude Code's 'don't ask again' feature exposed 2,539 secrets on GitHub, and AI Insider reports OpenAI hired Hacktron to probe its own systems.
What Comes Next: Mitigations and Industry Scrutiny
The researchers propose vendor-side cryptographic and system-level fixes to secure client-side reasoning, and offer guidance for users who may have exposed themselves to privacy risks.
HackerNoon's publication of the paper puts pressure on all three providers to rework how encrypted reasoning blocks are signed, scoped and validated across model families.
The disclosure also arrives during a broader week of AI security news tracked by AI Insider, including OpenAI's Hacktron partnership with expanded Daybreak Red frontier-model access, suggesting red-teaming and vulnerability research will intensify around agent infrastructure.
Sources
- hackernoon.com › Researchers Decrypt Hidden Chain-of-Thought From OpenAI, Anthropic and Google APIs
- opentools.ai › OpenAI, Google, Microsoft, and Anthropic Back Unified AI Training Opt-Out in Robots.txt
- theaiinsider.tech › The Week Ahead in AI: OpenAI Hires Hacktron for Probe, Anthropic Details Claude 'Dreaming', Post-Election US AI Safety, Plus Upcoming Earnings & Events