Research reveals vulnerabilities in reasoning blocks of Anthropic, OpenAI, and Google APIs

Google: Google develops and serves large language models such as the Gemini series through its cloud APIs, incorporating opaque reasoning blocks to enable replay without persistent server storage. The company focuses on scalable, secure AI infrastructure for broad application integration. Google's APIs were included in the study showing that reasoning traces could be ported across sessions and models within the same provider family. OpenAI: OpenAI builds and maintains advanced generative AI systems, including GPT models, and distributes them via public APIs that support features like encrypted reasoning for session continuity. Its APIs handle a wide range of enterprise and developer workloads that can involve sensitive data handling. The reported research specifically examined OpenAI's reasoning blocks as part of the attack surface allowing weaker models to decode traces from stronger siblings. Anthropic: Anthropic develops and provides frontier AI models through its Claude family, emphasizing safety and alignment in large language model systems. The company offers API access that includes opaque reasoning blocks for efficient multi-turn interactions without server-side storage of internal reasoning. In this news, Anthropic's API is identified as one of the providers whose reasoning trace mechanism was found vulnerable to cross-model reuse attacks. Research Focus: The paper demonstrates how encrypted reasoning fields in public agent logs can serve as an attack vector when compatible weaker models are used to decode them. Security Practice: Major LLM API providers return opaque reasoning blocks specifically to allow clients to replay them in later turns while avoiding server-side retention of the underlying reasoning. Vulnerability Disclosure: After the authors performed responsible disclosure, the affected providers implemented patches that rendered the reported cross-model reasoning trace attacks no longer reproducible.