Moonshot nerfs Kimi K3.1 Pro hacking skills ahead of open release
Moonshot deliberately weakened its new AI model's hacking skills because you cannot build a great coder without building a great hacker.
Models · Source: Reuters
What happened
Beijing startup Moonshot AI is releasing the weights for Kimi K3.1 Pro on November 4. Before putting the files online, the company deliberately crippled the model's offensive hacking abilities. They removed cyber tasks from its intermediate training and reinforcement learning environments. Reinforcement learning improves a model through repeated attempts at tasks and feedback. Moonshot removed this cyber material while keeping general coding work intact.
The lobotomy worked, but only slightly. On an internal software exploitation test called xdev-bench-hard, the model's score dropped from 90.5 percent to 81.0 percent. This test measures if a model can reproduce recently disclosed vulnerabilities in software like the Linux kernel and Node.js within three attempts. Even after the targeted cuts, the new model crushes its predecessor. Kimi K3 scored just 33.3 percent on the same test. The filtered K3.1 Pro retains a massive 47.7 percentage point advantage over the old version.
The timing is not a coincidence. Eight days before this announcement, China proposed new technology export controls. The Ministry of Commerce and Ministry of Science and Technology want to prohibit exporting AI that generates malicious code. They also want to restrict technology that uses AI to discover software vulnerabilities. Moonshot is getting ahead of the regulators by nerfing the model before release.
Key facts
- 81.0% — K3.1 Pro score on internal exploitation test after filtering
- 90.5% — K3.1 Pro score on internal exploitation test before filtering
- 33.3% — Predecessor Kimi K3 score on the same internal test
- November 4, 2026 — Date of planned open weights release
- Step 17 — Average progress Kimi K3 made out of 32 steps in The Last Ones network simulation
Why it matters
The open-weight release means anyone can download the model files and run them locally. Moonshot admits they will lose the ability to manage access or change safeguards once the files spread. This forces a hard choice for AI labs. You either release a powerful model that bad actors can use to automate network intrusions, or you release a weak model that nobody wants to use. Previous independent testing shows how fast these models are improving. The UK Artificial Intelligence Security Institute tested the older K3 model on a simulated corporate network called The Last Ones. K3 reached step 17 out of 32 attack steps on average. The new model is significantly smarter.
The deeper problem is that coding and hacking are the exact same skill. Moonshot kept general programming data in the training mix so the model remains useful for software engineers. But a model that can inspect code, follow a fault through several files, and test possible repairs can also find weak spots and break in. Good coding skills naturally transfer to offensive cyber work. An exploit is just a method of turning a software weakness into an action. You cannot build a great programmer without accidentally building a great hacker.
For builders
Local code inspection gets a massive upgrade
Builders can run Kimi K3.1 Pro locally to audit their own software. You can feed it older systems that receive little attention and have it hunt for vulnerabilities without sending proprietary code to an external API.
Fine-tuning can likely restore hacking abilities
Moonshot removed explicit cyber training data, but the underlying reasoning remains intact. Security researchers or attackers can likely fine-tune the open weights to restore the capability removed before release. Moonshot has not explained if this is possible.
Navigating new export control compliance
China is proposing bans on exporting technology that generates malicious code with AI. Founders building security tools on top of Chinese open-weight models must watch these regulations closely to avoid supply chain disruptions.
My take
You cannot separate coding from hacking. Moonshot is playing a PR game to appease Chinese regulators before dumping a highly capable model onto the internet. Once the weights are out, someone will fine-tune the safety rails away in a weekend. The real danger is not what the model knows today, but what anyone can teach it tomorrow.
Original reporting: Reuters. This is my rewrite and opinion.