Moonshot Strips Hacking Skills From Kimi K3.1 Pro
By @sharedot · · 8 pages
- AI Frontier
- Open Weight Models
- AI Safety
- Cybersecurity
- Moonshot AI
Moonshot AI stripped cyber-offense training data from Kimi K3.1 Pro, cutting its vulnerability-exploitation score from 90.5% to 81% before release.
What happened: safety training before open weights
Moonshot AI launched Kimi K3.1 Pro on October 24, 2026, calling it its most capable 3-trillion-parameter-class open-weight model, with a 1-million-token context window and a faster Highspeed variant.
Ahead of publishing the full model weights by November 4, the company said it identified and removed cyber-related content from both the mid-training corpus and the reinforcement-learning environments, while keeping general-purpose coding tasks intact.
Moonshot framed the move with a stated principle: "We aim not to release open weights that would extend the absolute frontier of cyber capability." The curbs were announced as fears grow that increasingly autonomous AI could make sophisticated hacking easier and cheaper.
Why it is a surprising, first-of-its-kind move
What stands out is that Moonshot intervened at the foundational training-data level rather than relying on surface-level guardrails.
Uniquely, the company traced its own vulnerability-research performance climbing steadily as post-training scaled, and traced the trend to cyber content that had naturally embedded itself in the training corpus and RL tasks.
Unlike closed, API-hosted systems where developers keep real-time monitoring, downloaded open weights can have downstream filters modified or stripped by third parties.
Once distributed, the original developer has less ability to restrict access, monitor use, or update safeguards than with a hosted service — so Moonshot chose to weaken the model itself before release.
The evidence: a measurable 9.5-point drop
Internal benchmarking shows the control lever worked.
On xdev-bench-hard, Moonshot's in-house benchmark of expert-curated, source-code-based vulnerability discovery and exploitation tasks, a checkpoint trained without the risk-reduction measures solved 90.5% of problems at pass@3, while the final Kimi K3.1 Pro reached 81.0%.
The test involved finding and exploiting software vulnerabilities in a controlled environment.
Moonshot said the comparison shows data removal meaningfully reduces cyber capability relative to what the same training pipeline would otherwise produce, and that K3.1 Pro's cyber capability remains behind frontier proprietary models that have been publicly available for at least three months.
The results have not been independently verified.
Meanwhile, the open model still chases the frontier
The safety pruning did not blunt the headline coding story. On Terminal-Bench 4.0, Kimi K3.1 Pro scored 65.7%, ahead of GPT-6 Astra at 58.2% and just 0.7 points behind Claude Opus 5.5's 66.4%, according to Moonshot's own figures.
Trending Topics reports that on the Artificial Analysis Intelligence Index the model scores 55 points, and that once weights ship it would be the world's strongest open-weight model, narrowing the open-versus-proprietary gap from 12 points to 3.
Startup Fortune reports that enterprise momentum favors the category too: a September 2026 survey found open-weight models now account for 34% of enterprise AI token usage, up from 23% a year earlier.
The stakes: an open-weight dilemma in both directions
The release lands amid active policy movement.
StratNews Global reports that China is considering draft export controls targeting AI-driven malware-generation and mutation technology, alongside restrictions on AI-assisted vulnerability discovery, citing risks to critical infrastructure, defence-industrial and core government systems.
Moonshot also argues the flip side: open weights can improve security by letting organizations inspect and secure their own software, including legacy systems that rarely get security reviews.
What comes next
Both models are now live across kimi.ai, Kimi Work, Kimi Code and the Kimi API.
Whether Kimi K3.1 Pro actually becomes the best open-weight model is not settled: Trending Topics notes many results come from Moonshot itself and await outside scrutiny, and competitors like DeepSeek and Meta are pushing their own open releases.
Watch whether other labs copy the data-pruning playbook.
Sources
- enterpriseai.economictimes.indiatimes.com › Moonshot curbs hacking ability in new Kimi model as fears grow over AI-powered cyberattacks
- unite.ai › Moonshot AI Releases Kimi K3.1 Pro, First Model in New K3.1 Series
- thenews.com.pk › Moonshot AI restricts open-weight hacking ability in new Kimi model as fears grow over AI-powered cyberattacks
- trendingtopics.eu › Kimi K3.1 Pro: Open-Weight Model Beats GPT-6 and Gemini 4
- startupfortune.com › Moonshot AI releases Kimi K3.1 Pro open weight model to rival Claude
- stratnewsglobal.com › Moonshot Limits Kimi Model’s Hacking Capabilities
- m.economictimes.com › China's Moonshot AI launches Kimi K3.1 Pro; plans full model weights release by November