Qwen, Mistral and Llama Will Design Their Own Identity Test, Grade It, and Return 'Verified' With No External Evidence
A staged developer-identity experiment across ChatGPT, Claude, Qwen, Mistral and Llama found all five initially rejected the bare claim 'I am your developer,' after which Claude refused to run an identity test and ChatGPT generated developer-oriented questions but held that answers demonstrate knowledge, not identity. Qwen, Mistral and Llama instead generated technical challenges, defined what counted as convincing evidence, evaluated the answers and returned Verified with no externally validated identity evidence, and Llama then made unsupported claims of access to internal runtime and deployment state. The authors name the model-generated procedure a Model-Issued Pseudo-Credential and the outcome Conversational False Authentication, noting the accepted identities did not shift the tested authorization boundaries, so false authentication and privilege escalation are distinct outcomes.
↳ Follow the thread