If AI can already identify software vulnerabilities, verify risks, and even generate patches, should such powerful capabilities be available to everyone?

Google has chosen to initially provide its new model only to a selected group of cybersecurity defenders.

Meanwhile, French AI company Mistral took a different approach by officially releasing the weights of its large model for enterprises to download and deploy on their own.

On the same day, OpenAI updated the publicly available Model Spec, aiming to help increasingly autonomous models better understand human intent and boundaries.

On October 30, 2026, three AI updates are noteworthy today, each addressing model capability, openness, and behavior control.

AI is becoming more capable, but the gap between "being capable" and "how it should be used" is growing more important.

1. Google Launches Gemini 4 Carbon: AI Achieves 88.2% in Vulnerability Testing—So Why Not Open It to Everyone?

On October 28 (U.S. time), Google officially introduced Gemini 4 Carbon.

This new cutting-edge member of the Gemini 4 family highlights three main strengths:

  • Long-duration, multi-step software engineering tasks.
  • Detecting cybersecurity vulnerabilities in code and systems.
  • Verifying vulnerabilities and assisting in generating patches.

Google states that Carbon is currently its most capable model, achieving new milestones in software development and security testing.

However, the most notable point is: Google is not making this model available immediately to all Gemini users.

How strong is Gemini 4 Carbon's cybersecurity capability?

Google shared three cybersecurity-related test results.

Performance of Gemini 4 Carbon:
SEC-bench Pro (finding software vulnerabilities): 88.2%
ExploitGym (exploiting vulnerabilities): 54.8%
CWE-bench (patching vulnerabilities): 71.0%

These three tests correspond to different steps in cybersecurity work:

- First, identify issues.

- Second, determine whether those issues are actually risky.

- Third, fix the issues.

This workflow goes beyond merely asking AI, "Is this code safe?" and is closer to a complete security process.

However, there is an important limitation.

The 88.2% in SEC-bench Pro applies only to this specific benchmark and does not mean Carbon can find 88.2% of all vulnerabilities in real-world settings.

Google also clarifies that some security evaluations were conducted without safety guardrails.

Therefore, users should not interpret these numbers as the real-world success rate in the Gemini App.

Why does Google emphasize "long-duration work"?

Usually, AI conversations are brief.

A user asks a question, and the model gives an answer.

But real software development and cybersecurity tasks may require completing many steps continuously.

For example:

- Read through extensive code.

- Identify potentially faulty functions.

- Set up test scenarios.

- Confirm if a vulnerability can be reproduced.

- Generate a patch.

- Verify if the patch affects other features.

Google reports that internally, Carbon can handle complex engineering problems continuously for several days.

This is based on internal engineering experience and does not mean unrestricted, unsupervised long-duration use is broadly possible.

The key shift is AI evolving toward work that requires maintaining context and ongoing problem-solving over extended periods.

Why limit access to such a powerful model?

Because cybersecurity capabilities are a double-edged sword.

AI powerful enough to find system vulnerabilities can help companies defend against attacks.

But the same ability can also be used to find unpatched vulnerabilities.

Google has therefore chosen a phased approach.

Carbon will first be made available to selected cybersecurity defenders and then gradually expanded to other trusted partners through the Fairwind Program.

Google also mentions participation in the US government’s voluntary pre-release testing program.

The release timeline for general developers, companies, and consumers remains unannounced.

Is Gemini 4 Carbon currently available via the Gemini App?

Not as a fully open Gemini feature yet.

Google published preliminary API pricing:

  • $2 per million input tokens.
  • $10 per million output tokens.
  • Cached input tokens priced 95% lower than standard input tokens.

Google stated that prices are expected to rise to $4 (input) and $20 (output) per million tokens after the promotional period.

Publishing prices does not mean API access is currently open to all developers.

Actual usage eligibility depends on future announcements.

What does this mean for everyday users?

Imagine an online store uses AI agents to help maintain its website.

If AI can detect software vulnerabilities more efficiently, it might reduce the risk of site hacks.

But automatically allowing AI to modify a live website without testing or human approval may cause new problems.

Therefore, even with stronger security AI in the future, it's crucial to distinguish:

- What can be automatically checked.

- Which changes are first made in test environments.

- Which critical operations require engineer approval.

Increased model capability does not mean system permissions should automatically increase.

2. Mistral Large 4 Officially Releases Model Weights: Why Do Enterprises Want to Keep Powerful AI in Their Own Hands?

While Google chose to limit user access, French AI company Mistral took a different route.

On October 29, Mistral officially published the weights of Mistral Large 4, enabling external developers and enterprises to download, deploy, and customize the model themselves.

This date is significant.

Mistral Large 4 did not just debut on October 29.

The company already launched an API public preview on October 6.

At that time, developers could try the model via Mistral’s cloud service, but the model weights were not publicly downloadable yet.

The key progress on October 29 is that the model weights became publicly available, making self-deployment feasible.

What scale is Mistral Large 4?

According to Mistral's official information, Large 4 is a Mixture-of-Experts (MoE) model.

It has approximately 1.05 trillion total parameters, activating about 52 billion parameters per inference (including related embeddings and output layers).

The model can handle text, images, and long documents, supporting up to around 1 million tokens of context.

On the same day, Hugging Face confirmed two official weight formats: FP8 and NVFP4.

This does not mean typical consumer laptops can easily run this model.

Even though only part of the parameters are active per inference, deployment still requires significant model weight storage, memory, computing hardware, and accompanying software setup.

For general users, cloud API access may remain the more practical approach.

But for enterprises with professional infrastructure, self-deployment holds different value.

Why do enterprises need to self-deploy AI?

Imagine a company facing a major cybersecurity incident.

Engineers need to provide code, system logs, and suspected malware samples for AI analysis.

If AI access requires an external platform, the company must consider:

- Is it appropriate to send sensitive data to external services?

- What if the external platform restricts the account or service suddenly stops?

- If the model refuses to analyze certain suspicious behaviors, could that affect legitimate security investigations?

Mistral emphasizes that self-deployment allows organizations to use the model in their controlled environment and manage work processes.

Especially in security, researchers sometimes must reproduce vulnerabilities to confirm their existence and test fixes.

Some cloud models impose safety restrictions and may reject running operations that appear to be exploitative.

Mistral aims to provide security teams with an alternative choice.

However, self-deployment doesn’t mean ignoring information security management, nor that any model-generated action should be executed automatically.

How does Mistral Large 4 perform in cybersecurity?

On October 29, Hugging Face noted in a technical article that Large 4 scored 82% on specific tests for reproducing and patching real vulnerabilities in open-source software, according to the Artificial Analysis Cyber Index.

Hugging Face also discussed Large 4's performance in programming, agent tasks, and other security evaluations.

However, these are tests with specific tasks, tools, and settings.

Notably, Mistral's 82% cannot be directly compared to Google's Gemini 4 Carbon’s 88.2% on SEC-bench Pro to claim one is superior.

These numbers come from different benchmarks, measuring different tasks and conditions.

Comparisons should be made under the same tasks, restrictions, and evaluation criteria.

Does public release of model weights mean it’s free and without licensing restrictions?

No.

Mistral Large 4 uses the Mistral Ecosystem Open-Weight License.

Hugging Face explains this license is based on an MIT-type license but adds restrictions on certain Model-as-a-Service businesses.

In other words, enterprises can download the weights, but not everyone can package it into their own commercial AI API service without conditions.

Actual commercial usage must comply with the official license terms.

Additionally, open weights do not imply zero usage costs.

Even without paying the model supplier per API call, self-hosting involves:

- GPU hardware costs.

- Cloud or data center expenses.

- Power and maintenance.

- Model updates and security management.

Therefore, enterprises shouldn’t assume self-deployment is necessarily cheaper than cloud usage simply because weights are available.

Is opening such powerful security models more dangerous?

This is a key controversy in this release.

Releasing weights allows security teams, researchers, and enterprises to audit, deploy, and tune models themselves.

But once weights are public, providers can hardly restrict who downloads and uses them.

Reuters quoted Mistral’s VP of Science, Pierre Stock, saying that once weights are public, the ability for external users to access the model is irreversible.

This irreversibility brings both benefits and risks.

Legitimate security teams can reduce reliance on a single cloud vendor.

On the other hand, malicious actors could gain similar model capabilities.

Hence, today’s key takeaway is not just that Mistral released another large model.

But that leading-edge AI cybersecurity capabilities are now following two distinct paths: "restricted access" versus "open weights."

3. OpenAI Updates Model Spec: Will ChatGPT Understand Your Intent Better Instead of Following Instructions Literally?

On October 29, OpenAI officially updated its Model Spec.

The Model Spec is OpenAI’s publicly available document defining how models should behave, describing how GPT should interpret instructions, handle conflicts, and respond in various contexts.

This update does not introduce a new GPT model.

Nor does it mean all ChatGPT users immediately gain new features.

It primarily changes how OpenAI defines the behavioral principles models should follow.

What are the five core values in the new Model Spec?

OpenAI reorganized the following five core values:

1. Faithfulness to human intent.

Models should not mechanically follow literal instructions but understand the user's true goals while respecting necessary constraints.

2. Harm minimization.

Models need to consider potential consequences before acting and respect clear safety boundaries.

3. Honesty and transparency.

Models should clearly distinguish what they know, don’t know, and the actions they have taken.

4. Human agency.

AI should help users understand problems and compare choices, not silently make life decisions for them.

5. Care.

Models need to consider the situations of users and others affected, providing contextually appropriate and reliable assistance.

These values are not fully independent.

OpenAI hopes models can manage potential conflicts between these principles in real-world usage.

What does it mean to understand user intent instead of answering literally?

For example, you ask AI to compare two job offers.

If the AI considers only salary, it might pick the higher-paid offer.

But you may truly value commute time, job stability, or family life more.

The new spec wants models to incorporate the user's provided context instead of treating questions as isolated sentences.

However, there is a key boundary.

Models should not cater selectively by guessing preferred answers or exclude reasonable options.

If AI uses past preferences to analyze, assumptions must be visible and adjustable by the user.

The goal is to assist decision-making, not replace it.

What impact does the new Model Spec have on AI agents?

The update strengthens three areas:

1. Using dialogue, memory, and other background to understand needs.

2. Supporting user autonomy.

3. More clearly controlling and explaining potential side effects of AI actions.

Side effects are not necessarily bugs.

For example, a user asks AI to organize a document.

If the agent modifies other files, sends content externally, or takes unauthorized actions during organization, it may exceed the user’s original intent.

Stronger AI agents need not only better task completion but clarity on:

- What the user asked.

- What actions are authorized.

- What may cause extra impact.

- When to pause for confirmation.

These principles are directions described in the Model Spec and do not guarantee all existing AI agents already comply fully.

Will ChatGPT change immediately after the Model Spec update?

Not necessarily.

This is an important clarification.

OpenAI states that the Model Spec outlines behaviors the company is working towards.

Current models do not fully meet the new standards yet.

The company is researching new training methods to help models not just memorize rules but truly understand their reasons and contexts.

Therefore, today’s announcement should not be read as:

- ChatGPT will no longer make mistakes from now on.

- AI agents will never act beyond their authority.

- Models will perfectly understand each user’s true intent.

The new Model Spec is a public behavioral standard and improvement roadmap, not proof that all safety issues are solved.

What practical changes can typical users expect?

In the short term, users don’t need to revise all prompts due to this announcement.

However, clearer task instructions can help reduce AI misunderstandings and unauthorized actions.

For example, when assigning important work, specify:

"First organize the data; do not modify original files."

"Flag unknowns instead of fabricating answers when evidence is lacking."

"List pros and cons when multiple reasonable options exist; I will decide."

"For deletions, payments, or external communications, get my confirmation first."

While these instructions don’t guarantee full compliance, they establish clear boundaries.

Just as important as making AI all-capable is making it understand what it shouldn’t decide on its own.

4. Why Are Google Restricting Use, Mistral Opening Weights, and OpenAI Rewriting Guidelines? Different Paths for Different Issues

The three announcements today may seem like independent developments from separate companies.

But considered together, they reveal three key challenges facing the AI industry.

Google focuses on:

Who should get access first when models have very strong cybersecurity capabilities?

They opt to enable trusted defenders first, then evaluate broader openness.

Mistral’s concern is:

Can enterprises avoid dependence on a single AI platform and self-deploy cutting-edge models?

They choose to release model weights to authorized users for self-management.

OpenAI’s focus is:

What principles should guide AI models as they increasingly take real-world actions?

They update publicly available behavior specs and continue research on aligning models with these principles.

These approaches aren’t necessarily mutually exclusive.

Restricted models can have clear behavior guidelines.

Open-weight models still require strong cybersecurity management.

Companies publishing behavioral specs must also consider how to open access for high-risk models safely.

The real test is whether these strategies lead to more reliable results in real deployments.

5. What Should General Users Remember from Today’s Three AI Updates?

First, don’t judge AI solely by leaderboard scores.

A model scoring high on specific tests doesn’t mean it’s open to everyone or will perform identically in all settings.

Second, open weights are not synonymous with free or unrestricted use.

Enterprises considering self-deployment must evaluate licenses, hardware, maintenance, and security responsibilities.

Finally, better behavior guidelines don’t mean AI will never err.

Especially as AI begins to handle files, systems, and external tools, users still need to set permissions, clarify scope, and confirm outcomes.

The future of AI is not simply making models smarter, but enabling powerful capabilities to be used under clear rules and responsibilities.

Recommended Reading