Moonshot AI restricts open-weight hacking ability in new Kimi model as fears grow over AI-powered cyberattacks

Moonshot AI has actively filtered and pruned cyber-related training data for its open-weight Kimi K3.1 Pro model to intentionally restrict its offensive hacking capabilities.

This proactive safety intervention from the Chinese artificial intelligence firm comes amid growing fears that autonomous AI could mount industry-wide anxieties over the rapid escalation of autonomous, AI-powered cyberattacks and the unique vulnerabilities associated with downloadable open-weight models.

Key details of Moonshot's safety strategy:

To curb everyday safety challenges, the AI startup has targeted easy access data removal in new advance models.

During the development of the Kimi K3.1 Pro model, Moonshot identified that cyber-related content and vulnerability-research tasks had naturally embedded themselves into mid-training corpora and reinforcement learning (RL) environments.

To mitigate risk, the company stripped out these materials from both training phases while keeping general-purpose coding features intact.

Measurable performance drop:

Internal benchmarking highlights the effectiveness of this control lever. A test version trained without these risk-reduction measures scored 90.5% on finding and exploiting software vulnerabilities in a controlled environment, whereas the publicly released Kimi K3.1 Pro dropped to 81.0%.

Open-weight dilemma:

Unlike closed, API-hosted systems where developers retain real-time monitoring and centralized control, open-weight models present unique governance hurdles.

Once model weights are publicly downloaded, downstream safety filters can theoretically be modified or stripped entirely by third parties.

Moonshot's approach attempts to tackle this risk at the foundational training level rather than relying solely on surface-level guardrails.

Governments, technology companies and security researchers are growing increasingly concerned that AI could make sophisticated hacking easier and cheaper.

Models that can write code, operate software tools and complete long, multi-step tasks may help users identify weaknesses in programs and automate parts of cyberattacks.

Moonshot AI said, "we aim not to release open weights that would extend the absolute frontier of cyber capability."