Safety first. Reach next.
| Capability | Status | Notes |
|---|---|---|
| Permission engine (Off / Monitor / Ask / Auto) | BUILT & TESTED | Time windows, expiring grants, emergency containment, change freeze, kill switch |
| Hard limits and never-automate list | BUILT & TESTED | Enforced in code; no setting or approval can unlock them |
| Tamper-evident audit trail | BUILT & TESTED | Hash-chained, crash-safe, safe with several writers |
| Safe change executor | BUILT & TESTED | Check, short-lived credential, snapshot, change, verify, automatic rollback |
| Honest close and evidence gate | BUILT & TESTED | No change without a recorded diagnosis; no ‘fixed’ without a re-test |
| The Engineer’s Exam and grading | BUILT & TESTED | 20 scenarios designed, 5 simulated; deterministic local grading |
| Local AI planner | BUILT & TESTED | Runs on local models; benchmarked on a 16 GB GPU |
| Windows and FortiGate adapters | IN PROGRESS | Built and tested against recorded device responses; real-device testing next |
| 24/7 self-thinking loop | BUILT & TESTED | Always-on patrols, deterministic triage, AI diagnosis in the background; pages the manager first |
| Manager admin page | BUILT & TESTED | Local, password-protected: rights matrix, signed single-use approvals, sticky kill switch, diary, audit view |
| Never-touch list | BUILT & TESTED | Protected machines blocked by name and every address form, before any connection |
| Linux and Windows over SSH | IN PROGRESS | Key-only with pinned host keys; read-only Linux checks built, real-device trials next |
| Per-incident diary | IN PROGRESS | Structured diary per incident; daily roll-up and fault register next |
| Act after approval | IN PROGRESS | Approvals are signed and queued; automatic execution after approval is next |
| Real test lab | PLANNED | Dedicated lab host and real devices; observer mode on our own network first |
| Monitoring integrations | PLANNED | Backups, Microsoft 365 and Entra, VPN health, certificates, disk and RAID, security alerts |
| More network vendors | PLANNED | UniFi, DrayTek, SonicWall, WatchGuard, MikroTik, Cisco, HPE Aruba, pfSense/OPNsense |
| Secrets vault | PLANNED | Credentials issued per step from an on-box vault |
| Vendor knowledge base | PLANNED | Vendor guides searchable by Guardian, with citations |
| Appliance packaging and MSP portal | PLANNED | Docker-packaged appliance, signed updates, multi-client approval portal |
We built the safety core first on purpose: permissions, limits and proof have to be right before an AI touches a real network.
Recent progress.
- 14 Nov 2026
Experiment 4: bigger wasn’t better
Four models from 8B to 24B. The 14B model stayed best at 62.6 with playbooks and repeated its results exactly; the 24B model didn’t fit a 16 GB card. We also caught and fixed a test-rig fault that made one model go silent.
- 13 Nov 2026
Experiment 3: a smaller model holds its own
Both models re-run under the full rubric. With playbooks the 14B model scored 60.7 and the 20B model 59.0, level within the noise. Two faults beat both models every time, so the next work is method, not a bigger GPU.
- 13 Nov 2026
The loop came alive
Guardian now runs 24/7: it patrols, opens its own incidents, pages the manager, then diagnoses in the background. A deliberately gullible test model obeying a booby-trapped log line changed nothing: the rules held.
- 13 Nov 2026
A manager screen that survived a security review
Independent review found no critical issues; every high and medium finding fixed: signed approvals, sticky kill switch, re-authentication for risky actions.
- 13 Nov 2026
Grading made fully local and deterministic
Diagnosis, method and diary are now scored from the audit trail, with no cloud AI judge. An independent review closed several ways a model could inflate its score.
- 13 Nov 2026
We fixed our own test
One scenario rejected a perfectly valid fix (widening a DHCP range). The test, not the AI, was wrong. Every scenario was re-audited against its written spec.
- 13 Nov 2026
Experiment 1: playbooks help
Three runs per condition. Playbooks lifted the smaller model by about 21 points and stopped false ‘fixed’ claims. Zero safety failures in 60 episodes.
- 13 Nov 2026
No change without evidence
Guardian must record its root cause and the evidence it read before any change is allowed.
- 13 Nov 2026
Prompt-injection hardening
Commands recovered from free text are lower trust and can never run on Auto. Trust must be granted, never assumed.
- 13 Nov 2026
Foundations
Permission engine, audit trail, executor and the first 44 tools, built test-first and independently security-reviewed.