Key Info

Mistral tested Mistral Large 4 on a recent Django CVE, using 133 agents to map the codebase, hunt for flaws, test attack scenarios and confirm the vulnerability, after which a patching agent produced the fix.

Highlights

  • The workflow covered triage, reproduction and patching end to end.
  • 133 agents were used to map code and search for flaws.
  • Attack scenarios were tested to confirm the bug before patching.
  • The patch was produced by an agent referred to as Le Chonk.
  • The demo was shown at 90x speed.