Ethereum co-founder Vitalik Buterin has supported Moonshot AI’s decision to limit advanced hacking capabilities in its Kimi K3.1 Pro model, warning that poorly handled AI security risks could lead to restrictive identity checks and widespread surveillance.
Summary
- Vitalik Buterin supports limiting dangerous AI capabilities without restricting access based on users’ identities.
- Moonshot’s filtered Kimi model scored 81%, compared with 90.5% for its unfiltered counterpart.
- Buterin warned that unchecked AI attacks could encourage mandatory KYC and intrusive surveillance.
- Moonshot plans to release Kimi K3.1 Pro’s open weights by November 4.
In an October 26 post on X, Buterin defended Moonshot’s decision to exclude advanced offensive cybersecurity material from the model’s training, arguing that developers need more time to strengthen software against increasingly capable AI attacks.
The comments followed Moonshot’s October 23 announcement of Kimi K3.1 Pro, which is already available through its applications and API. The Chinese startup intends to publish the model’s full weights by November 4.
Vitalik Buterin supports limiting dangerous AI capabilities
Vitalik Buterin’s position centers on the distinction between restricting what an AI model can do and deciding which people should be allowed to use it.
“Discriminating against use cases is MUCH less bad than discriminating against users,” he wrote, supporting Moonshot’s decision to leave advanced offensive cyber capabilities out of training.
He argued that the world needs time to prepare for powerful AI-driven attacks by strengthening existing infrastructure, reviewing older software and applying formal verification to newly developed code.
Formal verification involves mathematically checking whether software satisfies specified requirements. Buterin has repeatedly supported the approach as a means of finding security weaknesses before programs are deployed.
In September, he rejected claims that AI-powered hacking would eventually make cybersecurity unwinnable, arguing that automated verification could improve defenders’ ability to secure software.
His latest comments did not suggest that removing dangerous training material would prevent other developers from producing more capable models.
When X user LOID asked whether competing laboratories would eventually develop advanced offensive capabilities anyway, Buterin replied, “Yes of course they will.”
Moonshot reports measurable impact from Kimi security restrictions
Moonshot introduced Kimi K3.1 Pro as an upgraded open-weight model built on its existing Kimi K3 architecture.
According to the company’s technical announcement, developers removed selected cyberattack-related data from training and reinforcement learning while retaining general programming capabilities.
The decision reduced performance on Moonshot’s internal exploit benchmark. A version without the restrictions solved 90.5% of tasks, compared with 81% for the released model.
The difference amounts to 9.5 percentage points, although the company has not established how that reduction translates into real-world cyberattack prevention.
Moonshot reported that the previous Kimi K3 model scored 33.3% on the same assessment, indicating that the restricted successor still performed considerably better in its own testing.
On Terminal-Bench 4.0, the new model achieved 65.7%, compared with Moonshot’s reported scores of 66.4% for Claude Opus 5.5 and 58.2% for GPT-6 Astra.
Those performance figures come from Moonshot’s testing and have not been independently verified.
The company has released a Highspeed version that it says operates two to three times faster than the standard offering.
Buterin warns AI failures could lead to stricter access controls
Beyond the immediate cybersecurity concerns, Vitalik Buterin warned that governments and technology companies could respond to serious AI-related incidents by imposing restrictions on ordinary users.
He described a possible future involving locked-down models, mandatory Know Your Customer checks and intrusive surveillance.
Under such arrangements, access to advanced AI tools could depend on a person’s nationality, identity documents or other credentials.
Buterin warned that some users could become “second or third or fourth-class” participants based on their qualifications or citizenship.
His concerns are consistent with his recent experiments involving private AI access.
On October 16, he introduced a privacy-focused agent setup combining Tor routing, anonymous payments through zkAPI and local prompt rewriting to reduce the personal information exposed to external models.
The system was designed to allow users to access AI services while limiting the disclosure of their identities and activity.
AI cybersecurity remains a concern for blockchain developers
Independent assessments have previously examined the offensive capabilities of Moonshot’s models.
In July, the U.S. Center for AI Standards and Innovation and the United Kingdom’s AI Security Institute evaluated Kimi K3 against simulated cyberattacks.
The researchers found that the model performed below leading cyber-capable systems but could still attempt offensive operations despite its existing safeguards.
On one simulated corporate-network attack, Kimi K3 completed an average of 17 steps out of 32, compared with 28.5 for the most capable U.S. models tested.
The assessment covered the previous Kimi K3 release, not the newly introduced K3.1 Pro.
Previously, Buterin has proposed directing more AI-assisted development work toward security testing and mathematical verification.
In May, he identified applications for the approach across Ethereum’s zero-knowledge virtual machines, cryptographic proofs, consensus systems and quantum-resistant security research.
