Ethereum co-founder Vitalik Buterin has supported Moonshot AI’s decision to restrict advanced hacking capabilities in its Kimi K3.1 Pro model, while warning that poorly managed AI security risks could result in mandatory identity checks and widespread surveillance.
In an October 26 post on X, Buterin said limiting what a model could do was preferable to restricting access according to who a user was.
“Discriminating against use cases is MUCH less bad than discriminating against users,” he wrote, backing Moonshot’s decision to exclude advanced offensive cybersecurity material from the model’s training.
Buterin argued that developers needed time to improve software and infrastructure before increasingly capable AI systems could be used for attacks. He pointed to reviewing older software and using formal verification – the mathematical testing of whether code meets defined requirements – as ways to identify weaknesses before deployment.
He has previously promoted formal verification as a method of improving software security. In September, he rejected claims that AI-powered hacking would eventually make cybersecurity impossible to win, saying automated verification could strengthen defensive capabilities.
Buterin accepted that restricting training material would not stop other developers from creating models with more advanced offensive tools. Asked by X user LOID whether rival laboratories would eventually develop those capabilities, he replied: “Yes of course they will.”
Moonshot announced Kimi K3.1 Pro on October 23. The upgraded open-weight model is available through the company’s applications and API, with its full weights due to be released by November 4.
The Chinese startup says it removed selected cyberattack-related data during both training and reinforcement learning, while retaining the model’s wider programming abilities.
That decision affected its internal exploit benchmark. The unrestricted version completed 90.5% of tasks, compared with 81% for the released model – a difference of 9.5 percentage points. Moonshot has not established how that reduction relates to preventing real-world attacks.
The previous Kimi K3 scored 33.3% on the same test. On Terminal-Bench 4.0, Kimi K3.1 Pro recorded 65.7%, compared with Moonshot’s reported scores of 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra. Those comparisons have not been independently verified.
Moonshot has also released a Highspeed version, which it says runs two to three times faster than the standard model.
Buterin warned that serious AI incidents could prompt governments and technology companies to impose locked-down systems, mandatory Know Your Customer checks and intrusive surveillance. Access could then depend on nationality, identity documents or other credentials, leaving some people as “second or third or fourth-class” users.
On October 16, he introduced a privacy-focused agent system using Tor routing, anonymous payments through zkAPI and local prompt rewriting to limit the personal information sent to external models.
Independent research has also examined Moonshot’s cyber capabilities. In July, the U.S. Center for AI Standards and Innovation and the United Kingdom’s AI Security Institute tested Kimi K3 against simulated attacks. It completed an average of 17 of 32 steps in one corporate-network exercise, compared with 28.5 for the most capable U.S. models assessed. The study covered Kimi K3, not Kimi K3.1 Pro.
In May, Buterin proposed using more AI-assisted development for security testing and mathematical verification across Ethereum’s zero-knowledge virtual machines, cryptographic proofs, consensus systems and quantum-resistant security research.
