Mistral AI’s security demonstration leaves key evidence out of public view: the company showed Mistral Large 4 coordinating 133 agents on Django code, but did not identify the vulnerability or disclose the proposed repair.
The clip assigns code review and attack testing to agents, then shows Large 4 attempting a correction. Mistral presented the recording at 90x playback but gave no wall-clock duration, token expenditure or count of candidate issues.
Django’s Nov. 3 security release addressed six issues, rating one moderate and five low severity. Mistral did not say whether the demonstration used any of them. The clip does not establish whether the vulnerability was previously known or whether an outside party validated the repair.
The repair itself is unavailable for technical review. Mistral made Large 4’s weights available for local use on Oct. 29.
