Vitalik Buterin accepts Kimi 3.1 deliberately excluding top-tier offensive network capabilities from its training data
Vitalik Buterin said he accepts Kimi 3.1 deliberately excluding top-tier offensive network capabilities from its training data. He believes that distinguishing by use case is far better than distinguishing by user, and that the world needs time to harden legacy code and formally verify new code in order to cope with stronger cyberattacks.
When launching Kimi K3.1 Pro, Moonshot AI said it had removed network-related data from the training corpus and reinforcement learning environment while retaining general programming capabilities. The company said that in its internal exploit benchmark tests, the filtered model scored 81.0%, while the unfiltered version scored 90.5%. The model is planned to release open-source weights by November 4.
Original sources:
Title: Vitalik Buterin X post Publisher: vitalik.eth (@VitalikButerin) Publication date: 2026-10-25 URL: https://x.com/VitalikButerin/status/2114534202788225513 Source type: social media post Verification status: public content read in this session Supporting content: directly quoted his view on limiting Kimi 3.1's network capabilities
Title: Moonshot curbs hacking ability in new Kimi model Publisher: Reuters Publication date: 2026-10-24 URL: https://www.reuters.com/technology/moonshot-curbs-hacking-ability-new-kimi-model-fears-grow-over-ai-powered-2026-10-24/ Source type: news report Verification status: read in this session Supporting content: Moonshot removed network data, internal benchmark test results of 81.0% versus 90.5%, open-source weights plan
ABAB AI Insight
Moonshot chose to proactively reduce the model's offensive cyber capabilities before releasing open-source weights. Once open source, the company's ability to control how developers use the model declines, so it manages risk through data filtering rather than post-release restrictions. This approach shifts the capability boundary to the training stage.
Vitalik separates this from restrictions based on user identity. He pointed out that if early AI cyber risks are not handled properly, they could lead to highly locked-down models, KYC barriers, and access tiers based on nationality or credentials. Filtering by use case preserves the model's accessibility to more users.
Internal testing shows that the filtered model still significantly outperforms the previous version. The company also said it may gradually increase cyber capabilities in the open-source weights in the future to support defensive uses. The actual effect depends on how the community modifies and deploys the model after open sourcing.
ABAB News · Cognitive Law
Risk control for open-source models occurs more at the training-data stage than through post-deployment access control.
Capability restrictions based on use case differ from identity-based restrictions in terms of accessibility.
Proactively reducing offensive capabilities can buy time for defensive hardening, but it does not eliminate the subsequent model iteration path.