Key Notes
- Vitalik Buterin backs limits on offensive AI capabilities rather than restrictions based on users’ identities.
- Moonshot says the final model scored 81% on its internal exploit test, below a checkpoint without risk-reduction measures.
- The benchmark uses up to three attempts and does not measure private-key recovery or real-world attack rates.
Ethereum co-founder Vitalik Buterin has backed Moonshot AI’s decision to hold back the most advanced offensive cybersecurity capabilities in Kimi K3.1 Pro. He argues that limiting what an AI model can do is preferable to restricting who can use it.
In an October 26 post, Buterin said developers need time to strengthen systems, review older software and formally verify new code before more powerful AI-driven attacks arrive. Formal verification uses mathematical methods to check whether software satisfies specified properties.
Read Vitalik Buterin’s statement on AI cybersecurity limits
— vitalik.eth (@VitalikButerin) October 26, 2026
Restricting Capabilities Rather Than Users
Buterin’s concern extends beyond individual exploits. He warned that mishandling the first major wave of AI risk could lead to tightly controlled models, widespread identity checks and intrusive surveillance.
In that scenario, access could depend on a person’s credentials or nationality. He presented it as a possible outcome to avoid, rather than an announced policy or an inevitable result.
His endorsement concerns the decision to leave top-end attack capabilities out of training. It does not establish that Kimi is incapable of offensive security work, or that the wider threat has been resolved.
What Moonshot Changed in Kimi
Moonshot’s technical report says the company removed cyber-related material from mid-training data and filtered cybersecurity tasks from reinforcement-learning environments, while retaining general programming tasks.
On its internal xdev-bench-hard evaluation, the final K3.1 Pro scored 81.0%, compared with 90.5% for a checkpoint without those interventions. Both figures use pass@3, measuring success with up to three attempts per task. The 9.5-percentage-point gap is a benchmark comparison, not an estimate of real-world attacks prevented.
The test largely reproduces recently disclosed software vulnerabilities in sandboxed settings. These are company-reported results, not an independent assessment of every deployment. Moonshot also says stronger general coding skills still transfer to vulnerability research despite the filtering.
The company plans to release the full model weights by November 4. It says the intended open release should make defensive code auditing more widely available without setting a new frontier for offensive capabilities.
Software Defenders Face the Same Timing Problem
The need to give defenders time already affects cryptocurrency software. Bitcoin Core’s October 22 policy change allows faster releases for confirmed vulnerabilities that are easy to rediscover, including with AI. CoinScreamer’s earlier report explains how the process pairs fixes with public security advisories.
That policy concerns software fixes and disclosure timing. It does not demonstrate a break in Bitcoin’s underlying cryptography. Likewise, Moonshot’s exploit benchmark measures software attack tasks, rather than the ability to recover cryptocurrency private keys.
Disclaimer: CoinScreamer is an independent media brand owned and operated by NuvexMedia LLC, publishing news, research, and market insights on digital assets and related technologies. NuvexMedia LLC invests in and collaborates with companies across the digital asset, blockchain, and technology sectors. These relationships do not influence CoinScreamer’s editorial coverage, and the publication maintains full editorial independence to provide accurate, timely, and objective information. © 2025 NuvexMedia LLC. All rights reserved. This content is for informational purposes only and should not be considered legal, tax, investment, financial, or other professional advice.
