Chinese AI firm Moonshot said it removed offensive-cyber material from training Kimi K3.1 Pro, whose model weights are scheduled for release on November 4. On the company’s internal vulnerability-exploitation test, the filtered version scored 81.0%, compared with 90.5% for an otherwise comparable unfiltered model.
What is confirmed
The score gap matters for AI safety, but it does not prove the model is harmless. Reuters identifies these numbers as company-run tests, not independent certification. Open weights support experimentation and inspection, but also make it difficult to recall a model once copies have spread.
Why it matters—and what remains uncertain
Recheck the release plan and outside evaluations when the weights launch; do not turn benchmark percentages into safety guarantees.
Context worth knowing
Offensive-capability tests in controlled environments measure defined tasks, not every way a model could be used. Removing training material may reduce some behaviours without preventing users from adapting the system. A meaningful comparison needs methodology, limitations, external evaluation and a response plan if openly distributed weights create security problems.
What comes next
Editorial note: Figures and decisions reflect the cited reporting cutoffs. They may change; consult the latest official update before taking action.