AI providers and privacy
Rule and secret scanning run entirely on your host. An AI provider is needed only for agent work:
- SAST triage (
rules-ai) and deep analysis (rules-ai-harness). - Agent-assisted scans: web and API, penetration, cloud, Kubernetes and the mobile agent engine.
AI output is advisory. Triage never changes a finding’s severity or removes it from the register.
Connect a provider
Section titled “Connect a provider”Go to Settings > AI > AI Provider, choose a provider, enter the model and API key, and a base URL if the provider needs one. The key is sealed at rest and never shown again.
| Provider | Base URL |
|---|---|
| OpenAI | https://api.openai.com/v1 by default |
| Anthropic | https://api.anthropic.com/v1 by default |
| Google’s OpenAI-compatible endpoint by default | |
| Azure OpenAI | Required. Your deployment URL. |
| OpenRouter | https://openrouter.ai/api/v1 by default |
| OpenAI-compatible | Required. Any endpoint that implements OpenAI chat completions. |
Test Connection sends one short request and shows the provider’s answer. Clear Stored Provider removes the configuration.
Each scan sends at most 100 findings for AI review by default, with up to 4 review calls in parallel. Administrators can change these model limits under Settings > Scanning.
If no provider is configured, scans that need the agent are refused with a message pointing to Settings > AI.
What is sent
Section titled “What is sent”| Use | Leaves the instance |
|---|---|
| SAST triage | Each finding’s rule ID, file path, line and evidence snippet. Not whole files. |
| SAST deep analysis | The source the agent reads. The scan’s coverage ledger lists the files examined. |
| Agent-assisted scans | The agent’s conversation and tool results, including the evidence it reports. |
All of it is masked first when the privacy policy is on.
Privacy policy
Section titled “Privacy policy”Choose the mode under Settings > Privacy and Network > Privacy.
| Mode | Effect |
|---|---|
| Redact | Replaces matched values with type markers in masked downloads and share links. |
| Tokenise | Replaces matched values with reversible placeholders in AI requests, and masks downloads and share links. |
| Off | No masking. Masked downloads are labelled as not masked. |
The policy recognises email addresses, payment card numbers, IBANs, SWIFT and BIC codes, high-entropy secrets, values assigned to password, token or secret names, known credential formats (AWS, GitHub, GitLab, Slack, Google and JWT), PEM private keys, IP addresses, host names, file paths, values from the instance’s own environment, and your own regular expressions.
Host names are masked when they look internal. The public name of the asset a report covers is kept. You can choose to mask every host name.
Masking reduces what leaves the host. It cannot recognise a value that matches no known pattern, such as a business name in free text.
Who processes the data
Section titled “Who processes the data”The AI provider is yours. The instance calls the endpoint and key you configure, and no Cerberon service sits in that path. Cerberon does not receive, store or process the data you send to your provider. That transfer is governed by your agreement with the provider.