Researcher finds OpenAI model hidden inside Meta's Muse
Session logs from Meta's Muse assistant show one subagent routed to a model labeled muse-special, with signature details the researcher says match OpenAI's API.
A developer writing on the blog mouse.dev said Friday that session logs from Meta's Muse assistant showed one subagent routed to an unusual model. The subagent surfaced while Muse was building a website. The model was labeled "azure/muse-special," distinct from the roughly 15 variants of Meta's internal Avocado model that the post said handle nearly all other Muse traffic.
The anomalous session carried a signature tagged gpt_responses_v1, with an encrypted payload beginning "gAAAAA." The post described that format as OpenAI's own encryption standard. The session also used tool-call IDs in a different pattern than Avocado's: a "call_" prefix followed by 24 mixed-case characters, versus Avocado's 32-character hex IDs, according to the post.
Muse's agent runtime ships roughly 15 Avocado variants, the blog said, alongside models from rival providers: several Claude Opus and Sonnet versions, GPT-5.5 and 5.6 variants, and Kimi K3. It also includes dedicated integration code for Anthropic's API. API-key access to those integrations is restricted to an internal inference-proxy service, the post said.
The developer concluded muse-special is likely an OpenAI model served through Microsoft's Azure cloud, though wrote that its exact identity could not be confirmed. Meta has not commented on the finding, which reached the Hacker News front page Friday.