Skip to content
Back to blog AI-Powered Penetration Testing

Guard rails for the AI itself: controlling what your testing data touches_

· 3 min read

Guard rails for the AI itself: controlling what your testing data touches

Most organisations now have an AI policy, or are writing one. And most of those policies ask the same question about any AI-powered service: what happens to our data inside the AI?

For a penetration testing platform, that question matters more than usual. The data involved includes your system details, test credentials and, eventually, a list of your vulnerabilities.

Here’s how Block8.ai puts guard rails around the AI itself, and around the data it touches.

Guard rail 1: the AI runs where you expect it to

•  AI inference in Australia. For our Australian SaaS platform, engagement data is stored in Australia (Sydney), and AI inference runs in Australia.

•  No third-party LLM calls by default. Inference doesn’t leave the platform’s contracted provider.

•  Your choice of model. Anthropic’s Claude is the default, but if your AI policy requires a specific model family or provider, the engine can be configured to use what you’ve already approved.

Guard rail 2: every AI action is on the record

Every inference call is logged for audit, and every command the testing engine runs is logged too. Each AI-generated finding carries tamper-resistant provenance: the specific AI run, model version and timestamps behind it. Those logs aren’t kept indefinitely. They’re purged regularly and held for no more than 60 days. If your AI governance team has questions about what the AI did during your engagement, there’s a record to answer them.

Guard rail 3: your data is separated and protected

•  Isolation at the data layer. Each organisation’s data is isolated at the data layer, not just in the application, so a compromised application server can’t bypass it.

•  Credentials locked away. Test credentials are encrypted in a secrets vault before storage and never shown again in the interface.

•  Encryption everywhere. AES-256 at rest, TLS 1.2+ in transit, and application-layer encryption for stored secrets.

•  Controlled access. Multi-factor authentication, layered authorisation checks and short-lived signed links for report downloads.

Guard rail 4: we clean up after ourselves

Every test except internal testing runs in cloud testing containers. The testing environment is torn down after the report is generated, and idle containers shut down automatically.

Some systems can only be reached from inside your network. For that internal testing only, we use our Virtual Appliance. Its disk is fully encrypted, including while the appliance is switched off. AI inference still runs in the cloud, so the same AI guard rails apply. When you retire the appliance, a factory reset cryptographically wipes the encryption keys, engagement data and configuration.

Guard rail 5: independent oversight

These controls sit inside an information security management system that’s independently certified to ISO 27001:2022, covering the platform, our processes and our people.

What to ask any AI testing provider

Whether or not you choose Block8.ai, these are the questions your AI policy should be asking:

•  Where does AI inference happen, and can data cross regions?

•  Which AI models and providers touch our data, and can we choose?

•  Is every AI action logged, and how long are the logs kept?

•  How is our data separated from other customers’ data?

•  What happens to our data when the engagement ends?

A credible provider should answer every one of them clearly. We’re happy to go first.

What to do next

Want to walk your AI policy through our platform? Talk to an Expert.

Contact Block8.ai

Keep reading

Faster. Smarter. Simpler_

Block8.ai delivers AI-powered penetration testing, validated by CREST-certified experts, from $5K.

Start testing

AI-Powered Penetration Testing for Everyone_