AI investigates with
humans at the wheel

Testing runs the way an attacker would run it. Scope is agreed with you first, enforced outside the AI, and tested for bypass routes.

Astero's environment

Sydney, on Astero-operated hardware

AI agent Own virtual machine

Plans the investigation and adapts as it learns.

Testing tools Own virtual machine

Browser, shell and scripts the agent operates.

Network restrictions

Enforced outside the agent. Only agreed hosts and IP addresses can be reached, and the agent cannot change the list.

Agreed with you first

  • Targets
  • Accounts
  • Testing windows
  • Prohibited activities

Everything else

No connection can be made

Independent kill switch

Cuts all network access from the assessment virtual machine. The agent cannot restore it.

Nothing installed in your infrastructure

Astero operates the platform. You do not install software or connect internal systems.

Scope changes need your approval

Targets, accounts, testing windows and prohibited activities stay as agreed.

Tested for bypass routes

The controls cover browser, shell and script traffic, and stay in force even if the agent behaves unexpectedly.

From agreed scope to verified fixes

Agree the scope

Targets, accounts, testing windows and prohibited activities are set before testing starts.

Investigate

DiscoverExploreChainDemonstrate

Testing starts from the outside, without credentials, source code or a complete asset register. Authorised test accounts enable deeper testing of applications and APIs. Astero monitors the assessment and retains connection logs, agent activity records and evidence.

Review and report

Astero reviews every finding. Demonstrated exploitation is kept apart from potential consequences and untested areas, with coverage and limitations stated.

Verify fixes

Agreed fixes are retested.

Models and data

Plain answers for security reviews.

Which models?

Anthropic Claude models, used through Claude Code. They plan investigations, operate security tools and interpret results.

Where does it run?

In separate Linux virtual machines on Astero-operated hardware in Sydney. Inference goes through Anthropic's commercial API. Assessment files are stored locally with encryption at rest and restricted access.

What can the model see?

Application responses, screenshots and testing evidence. That may include customer information, test credentials and session tokens.

Is customer data used for training?

No. Astero does not train foundation models on customer data, and Anthropic's commercial API does not use inputs or outputs for training by default.

What else is connected?

By default, Anthropic's model API is the platform's only third-party service integration.

Can inference stay in Australia?

Yes. Amazon Bedrock is available for enterprise customers. Astero is also evaluating hardware to run fully self-hosted models.

Talk through your scope