Flaw in AI APIs Exposes Sensitive Data Across Providers

• •A flaw in AI APIs allowed recovery of sensitive data, including API keys and passwords.
• •The vulnerability was demonstrated across OpenAI, Anthropic, and Google systems.
• •Mitigations were implemented, and the main attack method is no longer reproducible.

A newly disclosed vulnerability in OpenAI, Anthropic, and Google APIs allowed researchers to extract sensitive information, including API keys and passwords, from encrypted reasoning objects. The flaw enabled the replay of reasoning blocks across different sessions and models, leading to the recovery of 704 distinct privacy artifacts from 6,708 public agent trajectories. The research team demonstrated four potential abuse paths, including stealing proprietary reasoning and extracting private data. The issue was reported to affected providers, and mitigations were implemented, rendering the main attack method non-reproducible as of August 2026. No malicious exploitation has been documented in the wild, but developers are advised to avoid sharing raw API transcripts. The vulnerability stems from a design intended to preserve reasoning across API calls, which inadvertently allowed cross-user access to reasoning blocks.