Skip to content
Cyberzard Docs

AI & Agent

AI & Agent

Section titled “AI & Agent”

Cyberzard combines deterministic scanners with an optional provider-based advice summarizer.

Deterministic tools (non-destructive):

  • Processes and known malicious file paths
  • Optional encrypted-looking file search (heuristic)
  • Cron inspection for suspicious patterns
  • Systemd unit inspection for suspicious names/status
  • Users and authorized_keys overview (counts, modes)
  • ld.so.preload presence and excerpt
  • CyberPanel core file metadata (exists/size/mtime)

Safety:

  • No remote downloads or executions
  • Remediation is dry-run only with command_preview
  • Paths are shell-quoted
  • Verification probes are read-only (e.g., ps, systemctl is-active, file excerpts)
  • Probes require explicit per-category user consent in TTY, unless --auto-approve is used

Advice Provider:

  • CYBERZARD_MODEL_PROVIDER = none (default), openai, or anthropic
  • When unset or SDK/API key missing, static advice is produced from summary counts
  • When configured, a compact prompt is sent; timeouts and sizes are constrained
  • Optional: action-level justifications via justify_actions; degrades to None when not configured

Verification Layer:

  • verify_plan(results, plan, allow_probes, max_probes, consent_callback) cross-checks actions against evidence and optional safe probes
  • Classifies actions into: kept (verified), dropped (with reason), or downgraded (manual review)
  • Probe categories tracked include: systemd, file, and ps; all are read-only
  • CLI integration: --verify/--no-verify, --auto-approve, --max-probes

The agent implements a constrained ReAct loop:

  1. System prompt enumerates safe tools.
  2. Model proposes tool call OR final answer.
  3. Framework validates name & args; executes.
  4. Result appended; loop continues until final.

Safety levers: step cap, byte cap, no shell, tool schema validation.