Security Analyser: overview
Once a day, you can have Capell’s security scan ask an AI model for a second opinion. The advice appears on the Security page marked as coming from the analyser. Each piece shows its severity, a short summary and any evidence the model cites. The built-in checks run whether or not you use it.
You choose the AI provider in AI Orchestrator. OpenAI, Anthropic and Gemini also need a key in the server’s configuration, which a developer sets. Ollama needs none. A global administrator can also get advice at any time with the Run scan now button on the Security page. Each time, the model gets one request, listing the Capell version, the names of the built-in checks, and installed packages with versions, update settings and trusted and enabled status. It adds any known security advisories for them and a few on or off settings such as debug mode. Anything that looks like a password or key is left out first.
The model cannot see your code, live pages, account details or other settings, and Capell keeps only its advice on the kinds of problem the built-in scan already knows. It makes no changes, and an answer with nothing flagged does not mean the site is sound. What you get is advice to read beside the built-in findings.
What it does
Section titled “What it does”Security Analyser adds a forward-looking review step to Capell’s security scan. When the scan runs (the scheduled scan, the Security page, or the scan command described in the README), the package sends a bounded summary of the installation’s posture to the AI model configured in AI Orchestrator and turns the model’s answer into advice-only findings. Core records them with the source analyser next to the built-in and advisory findings.
What reaches the model
Section titled “What reaches the model”Only a structured inventory, never your configuration values:
- the list of registered weakness ids the model may use, and the installed Capell version;
- each installed package’s name, version, trusted and enabled flags and auto-update policy;
- boolean posture flags only (for example whether debug mode or secure cookies are on); free-form values, anything whose name or value looks like a key, token, password, credential or connection string, and any value with embedded credentials are dropped before the prompt is built;
- advisory identifiers, affected and fixed versions.
The prompt is capped at a configurable byte budget (prompt_byte_budget, 48 KiB by default). Sections that do not fit are omitted rather than truncated mid-value.
What it never does
Section titled “What it never does”- It never changes anything. Analyser findings are advice only: Core marks them as not remediable, and the approved-fix automation ignores them.
- It never invents weaknesses. A finding must name a weakness id that is registered in Core; anything else is dropped and counted in the log.
- It never logs prompts or model output. A failed or malformed review is logged with the exception class only, and the scan records the analyser as failed so earlier analyser findings stay open instead of being resolved against an empty answer.
Turn it off or tune it
Section titled “Turn it off or tune it”Set capell-security-analyser.enabled to false to keep the package installed but skip the review, or run the scan with --no-analyser. max_findings, max_output_tokens, model and redaction_patterns are documented in the package config file. The review runs through AI Orchestrator’s module security-analyser and capability review-posture, so AI Orchestrator’s provider settings, rate limits and policy guardrails apply.
For developers: see the README.