Can AI become more powerful without letting its dangerous capabilities grow unchecked?

On October 25, 2026, two important AI news stories provide different perspectives on this question.

Chinese AI company Moonshot has launched the next-generation Kimi K3.1 Pro. Before releasing the model weights publicly, they proactively removed certain cybersecurity-related training data to reduce the model's capability to be used in cyberattacks.

Meanwhile, the UK government announced the establishment of an AI Biosecurity Lab aimed at harnessing advanced AI to detect biological threats earlier and accelerate vaccine development and other medical countermeasures.

One approach limits capabilities that could be misused, and the other uses powerful AI to protect humanity.

These two directions are becoming key topics in cutting-edge AI development.

1. Launch of Kimi K3.1 Pro: Why Did Moonshot Intentionally Reduce Its Cyberattack Ability Despite a More Powerful Model?

On October 24, China’s AI company Moonshot officially released Kimi K3.1 Pro.

This is the first model in the Kimi K3.1 series and currently Moonshot’s strongest open-weight model with 3 trillion parameters.

According to official information, the new model supports a context window of up to 1 million tokens, enhancing capabilities in software development, 3D modeling, financial research, and long-duration agent tasks.

The company also launched a Kimi K3.1 Pro Highspeed version, claiming 2 to 3 times the execution speed of the standard version.

However, the most notable aspect is not the model’s size or its high program evaluation scores.

Moonshot openly admitted:

They deliberately limited the cybersecurity attack capabilities of the model before releasing the weights.

Why Intentionally Weaken Certain AI Abilities?

Today’s cutting-edge AI models can read code, operate development tools, check system errors, and persistently attempt highly complex technical tasks.

These abilities are very useful for software engineers and cybersecurity professionals.

For example, they help identify software vulnerabilities, understand security issues in legacy systems, and organize necessary patches.

But these same abilities can be exploited to locate vulnerabilities and lower the technical barrier to launching cyberattacks.

This issue is especially complicated for open-weight models.

Once the full model weights are released, external developers can deploy, fine-tune, and retrain the model on their own hardware.

This openness encourages innovation but makes it harder for the original developers to control subsequent uses.

Moonshot chose to adjust the training process before releasing model weights.

How Did Moonshot Reduce Kimi’s Cyberattack Abilities?

According to Moonshot official technical articles, the team noticed that as the scale of later-stage training increased, the model’s performance on vulnerability research tasks also improved.

They found some cybersecurity-related content in the mid-stage training data, and certain security-related tasks were included in the reinforcement learning phase.

Moonshot made adjustments in two phases:

First, they identified and removed specific cybersecurity-related training data.

Second, they excluded some cybersecurity tasks from the reinforcement learning environment while retaining general-purpose programming training.

The goal was not to make Kimi lose its coding or debugging assistance capabilities entirely.

Rather, they aimed to prevent the model’s cyberattack ability from scaling proportionally with its general programming skills.

What Does the Drop from 90.5% to 81.0% Represent?

Moonshot used an internal test set called xdev-bench-hard to compare models with and without the cybersecurity risk reduction measures.

This test simulates a controlled environment for reproducing software vulnerabilities, evaluating the model’s ability to perform vulnerability exploitation tasks.

The scoring method is pass@3, allowing multiple attempts to assess success.

The official results show:

  • Model without data removal: 90.5%
  • Final Kimi K3.1 Pro with mitigation: 81.0%

The difference is 9.5 percentage points.

Moonshot believes this indicates that removing some cybersecurity-related data can reduce the model’s performance on designated attack tasks.

However, two clarifications are important.

First, these are Moonshot’s internal test results and have not been independently verified.

Second, 81.0% is not the real-world cyberattack success rate and does not imply the model lacks dangerous capabilities.

It only reflects the model’s performance under this specific test design and conditions.

Moonshot also noted the new model’s cybersecurity capabilities still exceed the previous Kimi K3 generation.

So this is not about making the model completely harmless but about controlling the growth of specific capabilities.

Is Kimi K3.1 Pro Available Now?

Yes, but it’s important to distinguish between product services and model weights.

According to the October 24 official announcement, Kimi K3.1 Pro and the Highspeed version are available through Kimi.com, Kimi Work, Kimi Code, and the Kimi API.

However, the full model weights have not yet been officially released.

Moonshot plans to release the weights no later than November 4, 2026.

In other words:

Using the official Kimi K3.1 Pro services now does not mean you can download the complete model weights and deploy them on your own server today.

Enterprises and developers must wait for the formal release of weights, licensing terms, and deployment documentation before deciding to adopt it.

What Is the Bigger Picture Here?

Historically, new model launches emphasize:

bigger parameter counts, stronger inference, better coding skills, higher benchmark scores.

This time Moonshot brought another issue to light.

If a model can become more dangerous, should developers proactively restrict certain capabilities before release?

This doesn’t mean all AI companies must adopt the same approach.

Removing cybersecurity data may affect some legitimate defensive uses, and open models can still be retrained after release.

But it shows that frontier AI competition is shifting from simply pursuing higher scores to deciding which capabilities should be openly available and which require additional management.

2. UK Invests £100 Million to Establish AI Biosecurity Lab: Can AI Help Prevent the Next Pandemic?

Also on October 24, the UK government announced the creation of the AI Biosecurity Lab, abbreviated as AIBL.

This is a new national-level AI biosecurity research program.

The UK government states the lab will integrate AI and life sciences to detect biological threats earlier and speed up research on vaccines, drugs, and other medical countermeasures.

The government also claims this will be the world's first institution dedicated solely to AI biosecurity.

However, this is the planned role; it does not mean the lab is fully established yet or has already prevented any outbreaks.

Why Does the UK Need an AI Biosecurity Lab?

AI-driven scientific research capabilities are advancing rapidly.

Researchers can use AI to analyze massive biological data sets, compare protein characteristics, identify potential drug candidates, and help design research protocols.

These abilities could accelerate some medical research efforts.

However, stronger capabilities may introduce new risks.

The British government points out advanced AI systems might lower the technical barriers for harmful biological activities.

Therefore, the nation needs not just stronger AI but also enhanced detection, prevention, and response capabilities for biological threats.

The goal of AIBL is to combine AI strengths with biosecurity expertise to build technologies that can be used defensively.

What Will AI Biosecurity Lab Do?

According to the UK government’s current plans, initial work will focus on two key areas:

First, earlier detection of emerging biological threats.

For example, assisting researchers in analyzing relevant data to identify anomalies or potential risks worthy of further investigation.

But AI’s analysis still requires professional verification and does not mean models alone can confirm pathogens or predict outbreaks.

Second, accelerating vaccine, treatment, and other medical countermeasure development.

AI will collaborate with biological research teams to shorten data analysis, candidate screening, and research design phases.

However, from AI-generated candidates to actual safe and effective vaccines or treatments involves extensive experimentation, validation, clinical trials, and regulatory processes.

The research aim is not equivalent to completed medical products.

How Much Is the UK Investing?

This announcement involves several figures that can be confusing.

First, the UK government is investing £100 million to establish the AI Biosecurity Lab.

Second, there is another planned £65 million for enhanced biosurveillance and medical countermeasures for specific pathogens.

Together, these form the £165 million National AI Biosecurity Initiative.

Additionally, OpenAI and Anthropic will provide the AI Biosecurity Lab with up to £200 million in model access, compute, and technical support.

It’s important to note:

The tech companies’ contributions are mostly access to models, computing resources, and commitments to support—not direct cash donations.

Thus, you cannot simply add the numbers to claim the government received that amount in cash.

The UK government also announced a broader AI investment and security plan exceeding £600 million covering cybersecurity, chip design, industry, and public services, not all of which is solely for this lab.

What Are OpenAI, Anthropic, and Google Providing?

According to the UK government:

OpenAI will provide £100 million worth of compute and model access and offer authorized access to cutting-edge models through the Rosalind Biodefense program.

Anthropic will provide up to £100 million in model access, trusted model availability, and specialized technical support.

Google DeepMind will participate in collaborative research and is evaluating trusted access to its advanced Gemini and dedicated scientific models.

UK life sciences AI company Basecamp Research will also contribute models, technical capabilities, and expertise.

The key point is not making all models openly available to everyone to study biological threats.

The government emphasizes trusted access and collaborative research mechanisms.

Is the AI Biosecurity Lab Operational?

As of the October 24 announcement, the lab is still a newly established entity with initial collaboration arrangements.

The UK plans for it to operate as an independent nonprofit advised by a scientific advisory committee.

Future cooperation will involve UK Health Security Agency, animal and plant health authorities, defense science organizations, and drug regulators.

However, it cannot yet be claimed that research facilities are fully established or that validated outbreak prevention results exist.

The real questions going forward are:

Can it transform AI model capabilities into reliable, verifiable biosecurity tools?

Can it speed up threat detection in practical research?

And can it accelerate medical countermeasures while meeting safety and regulatory demands?

Answers require ongoing studies and public evidence.

3. Why Should These Two Stories Be Considered Together?

Moonshot and the UK government appear to take different approaches.

Moonshot limits potentially misusable abilities before model release.

The UK invests resources to use cutting-edge AI to enhance biosecurity defenses.

But they face the same core question:

When AI has both beneficial and potentially dangerous capabilities, how should society decide who can use it and for what purposes?

In cybersecurity, AI can help analysts find vulnerabilities but also lower attack barriers.

In life sciences, AI can aid pandemic prevention and medical research but raise new biosecurity concerns.

This technology is typically called "dual-use technology."

Solutions are not limited to "fully open" or "completely banned."

Risk can be managed through layered measures.

Examples include evaluating capabilities before release, filtering specific training content, staged rollouts, trusted research access, and requiring independent validation of results.

Each measure has limits.

Training data filtering may not fully eliminate dangerous capabilities; trusted access requires rigorous review and continuous oversight.

Thus, assessing the effectiveness of safety measures requires transparency of testing, reproducibility of results, and continuous improvements.

What Should Everyday Users Know?

If you only use AI in daily life, you don’t need to immediately switch ChatGPT, Claude, or Kimi.

But start developing habits for evaluating new AI products.

When a company claims a safety measure is effective, first check if results come from internal tests or independent third parties.

When model weights are declared open, verify whether weights were truly released and review licensing and deployment conditions.

When governments announce major funding, distinguish between actual budgets, resource allocations, partnership commitments, and completed outcomes.

These details often determine whether an AI news story is an exciting vision or a verifiable advancement.

Trusting AI in the future relies not only on strength but on how its capabilities are verified, limited, and responsibly used.

Recommended Reading