Ethereum co-founder Vitalik Buterin publicly endorsed Beijing-based startup Moonshot AI on October 26, 2026, following the firm's decision to remove high-level offensive cyber capabilities from its new Kimi K3.1 Pro model.
Key Takeaways
- Moonshot AI removed cyber attack data from Kimi K3.1 Pro, reducing its exploit benchmark score from 90.5% to 81%.
- Open weights for Kimi K3.1 Pro are scheduled for public release on November 4.
- Binance Research discovered that AI currently exploits smart contracts roughly twice as effectively as it detects security flaws.
- Competitor DeepSeek's rival open model, V4.1-Flash, has already recorded an 85.7% score on the same exploit benchmark.
Open Weights and the Race for Cyber Defense
Moonshot AI plans to release the open weights for Kimi K3.1 Pro on November 4, allowing developers to run the model on independent hardware. To prevent misuse upon release, the startup excluded offensive cyber attack data from its training corpus and reinforcement learning environments, while preserving general software programming capabilities. The safeguard reduced the model's success rate on internal exploit benchmarks from 90.5% to 81%.
Vitalik Buterin argued that restricting functionality is significantly less harmful than restricting user access. He emphasized that cybersecurity teams need time to audit legacy code and harden defense mechanisms before automated exploits escalate. The necessity for defensive preparation is heightened by prior findings from Binance Research, which revealed that AI tools identify smart contract vulnerabilities for exploitation at twice the rate they assist in patching them. This risk intersects broader industry discussions regarding decentralization and open AI architectures.
Regulatory Surveillance Risks and Alternative Models
Buterin warned that if early releases of powerful open-weights models lead to major security failures, regulators could impose strict controls, including mandatory Know Your Customer (KYC) verification and invasive surveillance protocols for model access. Under such restrictions, end-user permissions could be stratified by credentials and nationality.
However, self-imposed guardrails may face practical limits. Buterin acknowledged that competing labs will continue developing offensive capabilities. Data from Moonshot AI shows that DeepSeek's open-weights model, V4.1-Flash, already achieved an 85.7% benchmark score without similar safety filtering.
Why It Matters
As open-weights artificial intelligence models approach parity with closed commercial systems, the window for securing blockchain smart contracts and legacy IT infrastructure is shrinking. Buterin's stance highlights a strategy of buying time for defensive hardening rather than relying on permanent technological bans. If developers fail to adopt proactive security frameworks before unfiltered open models proliferate, regulators may enforce identity verification on developer tooling, threatening user privacy across open-source ecosystems.
