OpenAI recently addressed a security vulnerability in its application programming interface that permitted the extraction of raw reasoning traces from frontier models, based on information published by Infosecurity Magazine.
The vulnerability, identified by the firm Xint, utilized a technique called compaction smuggling to bypass protections intended to hide the chain of thought behind model responses. While OpenAI typically provides summaries to API users to prevent model distillation, Xint discovered that the context compaction feature could be manipulated. By hiding reasoning within a compressed conversation, researchers successfully forced other models to reveal the hidden data.
Although OpenAI implemented mitigations in August to prevent the replay of encrypted reasoning, these measures did not account for compaction processes. Xint demonstrated the attack using GPT-5.6 Sol, GPT-4.1, and GPT-4o models. OpenAI confirmed the issue, deployed a fix on October 23, and provided a $6,000 bug bounty. This discovery highlights ongoing industry concerns regarding model distillation and the security of internal reasoning processes.
Previous reports from Anthropic and OpenAI have noted similar efforts by external actors to extract hidden data, underscoring the necessity for robust safeguards as frontier models become increasingly complex and widely deployed across various sectors.
Source: Infosecurity Magazine
