Agentic AI SOC — How SDP Automates L1 + L2 in Malaysia

L1 + L2 automation, L3 augmentation

SDP implements a three-tier automation model that matches how senior SOC engineers actually think about workload.

L1 — Automated triage (closed by AI)

The agent receives the raw alert from the ingestion pipeline. It correlates against prior incidents indexed in a sovereign vector store, classifies severity, validates against false-positive heuristics, and either closes the alert with a documented reason or escalates to L2. Every decision is captured to the AI audit log with token rollup, model version, and prompt hash.

Typical L1 outcomes: – False positive — closed. Benign PowerShell sessions matching prior false-positive patterns, expected admin activity, scheduled task noise. – Low severity — closed with note. Single-source anomalies with no lateral indicators. – Escalate to L2. Cross-host correlation, unknown binary execution, identity-anomaly signals.

L2 — Automated investigation (drafted by AI, approved by humans)

The agent pulls telemetry across the customer estate, builds the investigation narrative, attaches evidence (process trees, network flows, identity events, threat intel matches), and drafts the customer-facing ticket. A senior analyst reviews the AI-drafted packet and either approves it for customer delivery or expands the scope.

This is the productivity unlock. Senior analysts no longer write investigation reports from blank pages — they review, refine, and approve.

L3 — Human-augmented hunting (driven by humans, accelerated by AI)

Senior threat hunters drive ad-hoc investigations through the AI Threat Hunting chat interface. Reasoning-grade models walk lateral-movement chains, validate hypotheses against telemetry, and pivot between MITRE ATT&CK techniques. Humans steer; AI computes.

Learn more about AI Threat Hunting →

Per-call token rollup + prompt-hash drift detection

Every AI call is hashed by prompt template. SDP rolls up token usage per tenant, per model, per prompt template, per day. Two outcomes:

1. Cost transparency. Customers and Simply Data operators see exactly which prompts cost what. No mystery AI bill.
2. Prompt drift detection. If a prompt template is silently changed — by a developer, by a config drift, by an accidental commit — the hash changes and the rollup separates the new template from the old. This protects against the “we changed the prompt and now AI quality dropped 30%” failure mode that plagues unaudited AI estates.

For regulated industries — banking under BNM RMiT, insurance under MAS TRM-equivalent regimes, healthcare under PDPA — this audit depth is increasingly a procurement requirement, not a nice-to-have.

Why agentic (vs traditional SOC)

The economic case for Agentic AI SOC is straightforward.

Traditional SOC at scale: Headcount scales linearly with alert volume. A Malaysian MSSP onboarding three new banking customers needs to hire three more L1 analysts, then a shift lead, then a queue manager. Quality slips because new hires need 6-12 months to reach productive depth. Burnout rates in L1 roles exceed 35% annually across the industry.

Agentic AI SOC at scale: Compute scales with alert volume. New customers onboard in days, not quarters. L1 quality is consistent — the agent does not get tired at 03:00. Human analysts move up the value chain to L2 review and L3 hunting, which are intellectually rewarding and easier to retain.

For Malaysian customers facing BNM RMiT 14-hour incident reporting windows and PDPA breach notification obligations, the speed of Agentic AI SOC is not a luxury — it is the difference between a Saturday-morning notification and a Saturday-afternoon regulatory penalty.

Risk + safety posture

Agentic AI in a SOC is not risk-free. Simply Data has engineered specific guardrails into SDP:

• No autonomous customer-facing action. AI agents close internal triage decisions. AI agents do not push response actions (hash block, isolate, kill) to customer endpoints without operator sign-off in the standard flow. Customers can opt-in to faster autonomous flows for specific high-confidence patterns, with full audit.
• PII redaction at the boundary. Optional in-line redaction of NRIC, passport, IBAN, and email addresses before content leaves Simply Data infrastructure for AI inference. Opt-in per tenant; opt-out where regulatory or operational context requires the unredacted prompt.
• Model versioning + prompt versioning. SDP captures the model version and the prompt hash for every AI call. When a model is updated or a prompt is edited, the change is visible in the audit log. There is no silent model upgrade that changes triage behaviour without forensic visibility.
• Silent-drop monitoring. A counter on the platform health endpoint surfaces audit pipeline failures (e.g. an AI call that completed but failed to write its audit row). Operators see the counter; an unexplained spike triggers investigation.
• Human-in-loop for L2 and L3. The standard flow keeps a human in the loop for all customer-facing outputs. Autonomous closure is restricted to internal triage decisions with high-confidence patterns.

These are the controls a regulator is going to ask about when they audit an MSSP using AI in incident response. SDP is engineered to give clean answers.