AI for business

GPT-6 Astra and the new stage of AI: more capability needs better controls

OpenAI presents GPT-6 Astra and reports a Critical cybersecurity threshold in its own evaluation. For businesses, permissions, context, review and traceability matter most.

9 September 2026 · Updated on 2 November 2026 · 7 min read · My NovaGen AI team

OpenAI presents GPT-6 Astra as its new model. According to press coverage, access began with a limited group of organisations and was then extended to paid plans; according to some reports, the API came later, where the model appears as gpt-6-astra.

One detail stands out: in its safety evaluation, OpenAI says Astra reaches the Critical threshold of its Preparedness Framework in cybersecurity. It also describes safeguards, initial restrictions and plans to expand defensive access. These are the provider's own claims.

That does not mean any company will hand control of its systems to an autonomous model tomorrow.

But it is a trend worth noting:

models are increasingly able to act, not just answer.

And when an AI can do more, the system around it matters even more.


The more capable the models, the more value there is in defining permissions, context and checkpoints before actions run.
The more capable the models, the more value there is in defining permissions, context and checkpoints before actions run.

From an assistant that drafts to a system that acts

In the first years of business adoption, most AI use was assistive:

  • drafting;
  • summarising;
  • brainstorming;
  • translating;
  • analysing documents.

The user asked for something, got an output and decided what to do next.

Now a different kind of use is spreading:

  • querying tools;
  • navigating systems;
  • running tasks;
  • updating records;
  • coordinating workflow steps.

1. More intelligence does not remove the need for architecture

It is easy to assume a more capable model needs fewer rules.

In reality, the opposite can be true.

If an AI only drafts text, the risk is limited to what it writes.

If it can:

  • open applications;
  • query data;
  • change records;
  • run code;
  • send messages;
  • interact with external systems;

then we need to define far more precisely what it is allowed to do.

A model's capability and a system's permissions are two different things.


2. The model should not get unlimited access just because it can

A basic principle is least privilege: give only the access that is necessary.

If an agent needs to check a calendar, it does not need access to the whole infrastructure.

If it needs to create a task, it should not be able to delete projects.

If it needs to answer a customer question, it has no reason to read internal financial data.

The architecture should restrict:

  • tools it can use;
  • operations it is allowed to perform;
  • records it can access;
  • actions that need confirmation;
  • environments where it can act.

The more capable the model, the more the environment must set the boundaries.


The competitive edge increasingly lies in the system around the model: data, integrations, evaluation and traceability.
The competitive edge increasingly lies in the system around the model: data, integrations, evaluation and traceability.

3. Cybersecurity is no longer separate from AI

OpenAI says Astra reaches its Critical cyber capability threshold and that it has strengthened protections around the model.

The claim comes from OpenAI's own evaluation, not from an independent certification.

Even so, it matters, because it shows how fast model capabilities are moving.

For a company, this reinforces several practices:

  • keeping secrets out of instructions;
  • protecting tokens and credentials;
  • separating environments;
  • limiting permissions;
  • logging actions;
  • reviewing integrations;
  • requiring confirmation for sensitive operations.

Security can no longer be "IT's job" and something to check afterwards.

It is part of the agent's design.


4. A useful agent needs context, not blanket access

Business systems work best when the AI receives only the context the task needs.

Not when it gets the entire database.

For example, to answer a customer about an appointment, it may only need:

  • the service requested;
  • availability;
  • the customer's name;
  • the conversation ID.

It does not need the full billing history, private notes or information about other customers.

This improves privacy, security and operational clarity.


5. Human review shifts to the critical points

With more capable agents, the goal is not to check every sentence by hand.

Human review can focus on the critical points.

For example:

Before acting

  • sending a quote;
  • deleting a record;
  • making a purchase;
  • changing commercial terms.

When confidence is low

  • ambiguous intent;
  • contradictory data;
  • missing information.

When the user asks for it

  • the customer asks to speak to a person;
  • a complaint;
  • an exception.

This lets you automate volume without losing control where it really counts.


6. Agents need continuous evaluation

This matches what we see in every real-world automation project.

Testing on launch day is not enough.

You need to review:

  • changes to the model;
  • changes to the knowledge base;
  • new exceptions;
  • errors that build up;
  • actions that should never have run;
  • behaviour with unexpected inputs.

The agent also needs to change when the business does.


7. The big mistake would be selling "autonomy" as the absence of control

The word "autonomous" gets used a lot in tech marketing.

But useful autonomy does not mean total freedom.

A well-designed autonomous system can include:

  • a clear goal;
  • limited tools;
  • a spending cap;
  • banned actions;
  • confirmation steps;
  • logging;
  • stop conditions.

In fact, those limits are what let you grant more autonomy safely.


The more capable a model is, the more security, permissions and technical architecture that limit its actions matter.
The more capable a model is, the more security, permissions and technical architecture that limit its actions matter.

8. What this means for a small business

A small business does not need to rebuild its whole infrastructure because a new model comes out.

The best strategy is still to start with specific processes.

For example:

  • routine customer service;
  • sorting requests;
  • preparing follow-ups;
  • controlled CRM updates;
  • document support;
  • internal analysis.

Increase autonomy only when:

  1. the process is well understood;
  2. the action can be measured;
  3. limits are in place;
  4. there is a clear audit trail;
  5. we know what to do if it fails.

9. The model stops being the only competitive edge

If several providers can access very powerful models, the difference shifts elsewhere.

To the system.

  • business knowledge;
  • well-structured data;
  • good integrations;
  • permissions;
  • testing;
  • user experience;
  • maintenance.

That is why we think "which model do you use?" matters less than another question:

"how is the system that uses that model designed?"


10. Our take for My NovaGen AI

The news does not push us to put AI everywhere.

It pushes us to double down on the approach we already believe in:

AI where it adds value, rules where they are enough, and human control where the impact demands it.

A more capable agent can greatly improve a solution.

But it can also amplify a bad permission, a bad rule or a poorly designed integration.

Capability and control have to grow together.


Practical summary: capability and control must grow together

In practice: a more capable model is no reason to give it more permissions by default. The system should grant only the context and tools each task needs.

If the agent can...What to control...
Query informationWhich sources and records it can read.
Modify dataWhich fields it can change, and what validation applies.
Send or execute actionsWhich actions need human confirmation.
Use external toolsCredentials, scope, traceability and limits.
Handle exceptionsWhen it must stop and hand over to a person.

Frequently asked questions

Does a small business need to switch models every time a more powerful one appears?

No. The value usually lies more in the process, the data, the integrations, the permissions and the testing than in constantly chasing the latest model.

Does more autonomy mean less human intervention?

Not necessarily. Useful autonomy rests on clear limits and keeps human involvement for sensitive decisions, exceptions and low-confidence situations.

What should I review before giving tools to an agent?

The purpose of the task, the data it needs, the operations it is allowed to perform, the actions it must not take, the confirmations required and the record it leaves behind.

To go deeper into process design, you can read everything that goes into an automation that actually works and the solutions page.

Conclusion

According to OpenAI, GPT-6 Astra reaches the Critical threshold in cybersecurity in its own evaluation. These are the provider's own claims, not verified by a third party.

But for a company, the real news is not only that a model can do more.

It is that we are entering a stage where AI can take a more direct part in doing the work.

That makes the following even more important:

permissions + context + evaluation + traceability + human intervention.

Models will keep changing.

A good business system must be able to benefit from those improvements without losing control each time the technology changes.


Source and review

Information taken from press coverage of the GPT-6 Astra launch. We could not check it against OpenAI's own publications. The figures and evaluations it describes are OpenAI's own claims, based on its own configurations. We have not verified them independently.

Share article

Does any of this sound like your situation?

We can review your specific case and tell you, with no commitment, what is worth automating and what is not.

Talk to Nova on WhatsApp