Researchers Claim Method to Extract Hidden Chain-of-Thought Reasoning from OpenAI and Anthropic APIs
Summary
- • Researchers claim frontier AI models' hidden chain-of-thought reasoning can be fully extracted via a 'deep_think' tool API trick.
- • When given a tool named 'deep_think' instead of native thinking mode, models call it outputting their full internal CoT in tool-call format.
- • Extracted reasoning token counts matched billed API thinking tokens 1:1, suggesting complete rather than partial exposure.
- • The claimed vulnerability reportedly affects APIs of every major frontier AI company; no official company responses issued as of August 12, 2026.
Details
Hidden CoT extraction claimed via tool trick
Researchers assert that giving frontier AI models a tool named 'deep_think' — rather than enabling native thinking — causes models to emit their full internal chain-of-thought reasoning as tool call arguments, bypassing output suppression.
Token count matches billed thinking tokens 1:1
Claimants verified that the volume of reasoning tokens extracted via the exploit matched the thinking tokens billed by the API for most prompts tested, implying full internal reasoning is exposed.
Claimed to affect 'every frontier AI company'
OpenAI and Anthropic are specifically named in the title and discussion; the broader claim is that this is a systemic API design vulnerability, not an isolated model-specific flaw.
Finding shared via social media, Aug 11, 2026
The claim originated in a tweet on August 11, 2026, which accumulated 744,000 views. A separate post in the thread states the team 'can finally talk about it' suggesting prior responsible disclosure or embargo period. No formal research paper published as of August 12, 2026.
Single low-trust source; unverified as of publication
The sole source is a social media post aggregated by Hacker News (trust score 0.5). No confirmation from OpenAI, Anthropic, or independent security researchers has been published. This event should be treated as an unverified claim pending further evidence.
Why hidden CoT matters
AI companies suppress visible CoT output as a design choice to protect model integrity, prevent adversarial manipulation of reasoning, and protect commercially sensitive internal logic. Exposed reasoning could enable prompt injection attacks targeting the reasoning process or reveal proprietary training artifacts.
Potential impact on AI safety research
Ability to read hidden reasoning could also be used constructively — for AI auditing, alignment research, or verifying that models reason as claimed. The same technique that is a security risk for operators could be valuable for independent safety researchers.
Source: Social media post circulated via Hacker News, August 11, 2026. Claim unverified by independent researchers or affected companies as of August 12, 2026.
What This Means
Frontier AI models from OpenAI, Anthropic, and reportedly other major labs suppress their chain-of-thought reasoning from API outputs — a design choice meant to protect both safety integrity and competitive IP. Researchers now claim this suppression can be bypassed by providing models a tool named "deep_think," causing them to emit full internal reasoning as tool call outputs, with token counts matching billed thinking tokens 1:1. If confirmed across multiple labs, this would represent a systemic API design vulnerability rather than a single-vendor bug, requiring coordinated industry fixes. The claim remains unverified by formal research or company acknowledgment as of August 12, 2026, but has spread rapidly in the AI research community and warrants close monitoring.