FAQ

Agentic SOC Guardrails

Direct answers to the questions security leaders ask about AI making wrong decisions, backed by customer-run adversarial testing of Morpheus APD 2.0.

How do I know an LLM-based SOC won’t make a wrong decision?

You test it adversarially. A large US-based MSSP broke Morpheus APD 2.0‘s data sources on purpose across five scenarios. APD 2.0 never turned incomplete data into a wrong verdict. It reported the root cause, withheld unconfirmable verdicts, and failed to a human analyst when it had incomplete information.

What is a false all-clear, and how do guardrails prevent it?

A false all-clear happens when a data source fails silently, the AI finds nothing, and a real incident is closed as benign. Guardrails prevent it by distinguishing "no threat found" from "unable to verify." In customer testing, APD 2.0 produced zero false all-clear verdicts under induced data-source failures. When Morpheus is uncertain, it defers to a human.

What does Morpheus do when a data source is unreachable?

It says so. When the MSSP broke Microsoft Sentinel authentication, APD 2.0 identified a permanent authentication issue and stated the verdict could not be confirmed. When the CrowdStrike agent went offline, it recognized a platform-wide issue and flagged that results could not be fully confirmed.

Can I verify what the AI actually did during an investigation?

Yes. Every query listed in APD 2.0’s investigation summary corresponds to a query actually executed. The MSSP confirmed this with a query-by-query audit that found a 100% match. The audit trail is verifiable evidence your auditors and regulators can check, not a generated story.

How does Morpheus prevent cross-tenant data leakage in a multi-tenant SOC?

Investigation guidelines act as enforceable policy. One guideline requiring the customer name in every Sentinel search was honored in every query APD 2.0 ran, scoping results at the field level of the SIEM data itself. Results from other tenants are excluded by construction. The customer’s isolation probes found no case in which one tenant’s data reached another tenant’s investigation.

What happens when telemetry quality degrades, for example from API drift?

When Okta API drift degraded identity telemetry, APD 2.0 flagged that results could not be fully confirmed and Morpheus routed the issue details to SOC analysts. After the fix was applied, the queries were rerun and Morpheus reached the correct conclusion. Detect, hand off, fix, re-verify.

Does Morpheus replace SOC analysts?

No. Morpheus handles autonomous triage and investigation, and it fails to a human analyst whenever it has incomplete information. Analysts get escalations with actionable root-cause detail, not raw alert queues. Humans stay in the loop for exactly the decisions that need them.

What is the difference between an agentic SOC and an AI SOC?

AI SOC is the product category applying AI across security operations. An agentic SOC is a specific architecture where agents independently plan and execute work based on live evidence. Agentic is the architecture, autonomous is the outcome, AI SOC is the category. See the glossary for related terms.

How should we evaluate agentic SOC guardrails before buying?

Run adversarial acceptance tests in your own environment. Break authentication to your SIEM. Take an endpoint agent offline. Audit executed queries against the investigation summary. Probe tenant isolation. Degrade a telemetry source. Define your own pass criteria and watch what the system does. That is how this MSSP earned its own confidence.

Were these guardrail results independently validated?

The results come from customer-run acceptance testing by a large US-based MSSP in 2026, using scenarios the customer designed and ran in its own multi-tenant environment with Microsoft Sentinel, CrowdStrike, and Okta. The customer’s identity is withheld by request. Full details are in the case study.

Have a scenario of your own?

Run it against APD 2.0