Business Continuity in the Age of AI
Organisations run their operations on entityOS.cloud. That makes the platform’s availability and integrity a continuity concern, not only a security one — and advanced AI is compressing the time between a new weakness and its exploitation. The staged “closed internet” model is built so the service stays reachable by you, and only you, even under pressure.
The things that make a platform valuable to run on also make it worth attacking. Advanced AI does not invent new targets — it makes reaching them faster, cheaper and more automated, shrinking the window a defender has to react.
Weaknesses are characterised and weaponised in hours, not weeks — the react-and-patch window narrows.
Probing that once needed a skilled operator now runs continuously and cheaply against every exposed endpoint.
Leaked keys, tokens and passwords are harvested and reused the moment they surface anywhere public.
A public endpoint is reachable by every attacker, all the time — exposure is the default, not the exception.
The platform an organisation runs on is a single, high-value place to aim at — one target, wide impact.
AI makes phishing and social engineering harder to spot, so “who is really connecting” matters more than ever.
Continuity means the platform is reachable by you, never by the internet.
Running on a shared cloud platform concentrates two things an organisation cannot afford to lose. Continuity planning treats both as first-class:
The public internet is treated as hostile. Access is earned in layers, so losing any one layer degrades the service rather than stopping it — and never hands control to the internet at large.
Each stage of the closed-internet model is usually described as security. Read another way, each one is also a continuity control: it narrows how the platform can be taken down or taken over, and gives you an independent path to keep operating.
When a weakness is suspected or confirmed and a fix is not yet available, these shrink exposure while the layered gates keep the service reachable by you.
Cut the Stage 1 allowlist to the minimum addresses essential operations need; remove everything else.
Require a valid client certificate for protected surfaces, so a network foothold alone cannot reach them.
Where credential integrity is in doubt, lean on verifiable identity so being “inside the network” confers no authority by itself.
Keep at least one rehearsed access route that never depends on the mechanism currently under stress.
Rotate credentials, tokens and certificates that could have leaked; assume anything exposed is untrustworthy until proven otherwise.
Detection draws on telemetry independent of the affected component — never only on the logs it controls.
Residual ratings assume the staged-access controls are in place; before those controls, every rating is higher. Owner: Security & Platform Engineering, escalating to the executive sponsor on activation.
| ID | Risk | Likelihood | Impact | Residual | Primary treatment |
|---|---|---|---|---|---|
| C1 | Platform unreachable (volumetric attack or upstream outage) | Possible | Major | Medium | Allowlisted edge; diverse alternate path; tested failover |
| C2 | Credential or token theft leads to account takeover | Likely | Severe | High | mTLS per-connection identity; short-lived secrets; rotation |
| C3 | AI-accelerated intrusion outpaces manual response | Likely | Major | High | Layered gates; independent detection; automated containment |
| C4 | A single perimeter mechanism is bypassed by a zero-day | Possible | Severe | High | Diverse gates; no single mechanism is trusted alone |
| C5 | Data or configuration integrity silently altered | Possible | Severe | Medium | Identity-anchored access; out-of-band, append-only logging |
| C6 | No vendor fix available for an extended window | Likely | Moderate | Medium | Virtual patching; tighten allowlists; elevate to mTLS-only |
Six representative risks shown; the full register is maintained in the ISMS.
Each layer enforces something different, in different code, at a different level. Defeating one does not hand an attacker a certificate, a credential, or a place on the allowlist — so the platform stays reachable by you.
The property that matters is diversity of defence: the layers do not share a vulnerable component, so no single weakness opens all of them at once.
Continuity comes from rehearsed alternate paths and drills, not from a document. The organisation keeps at least one access path that never depends on the mechanism under stress.
A second route in a separate zone, terminating into an mTLS-gated entry — the continuity spine that keeps essential services running.
Pre-authorised, time-boxed emergency access, guarded by identity and logged out-of-band. Never a shared key, never the affected component.
Tier-0 systems: RTO ≤ 1 hour, RPO ≤ 5 min. Essential systems degrade but stay available; standard work defers until recovery.
Bypass and takeover alerts come from out-of-band, append-only telemetry — sources the affected component cannot alter.
Where persistence is possible, restore from a known-good image and re-attest integrity before returning a component to service.
Quarterly failover and break-glass rehearsals; half-yearly zero-day tabletops. A capability never exercised is an assumption, not a control.
In the age of advanced AI, the platform an organisation runs on must stay available and trustworthy under pressure — not just secure on a good day.
Diverse, independent gates turn “expose or halt” into “degrade and continue”. That is the continuity value of the closed-internet staged model.