---
title: What does Agent Engine Threat Detection in Security Command Center detect?
url: https://opencontroller.lyzr.com/blog/vertex-ai-agent-engine-threat-detection
published: 2026-11-17
updated: 2026-11-17
author: Lyzr
---

# What does Agent Engine Threat Detection in Security Command Center detect?

**Agent Engine Threat Detection, now documented by Google as Agent Platform Threat Detection, is a preview Security Command Center service that watches AI agents deployed to Vertex AI Agent Engine's runtime. It flags infrastructure-level attacks such as reverse shells, malicious binaries and container escapes, plus control-plane signals such as agent identities probing the metadata service or exfiltrating BigQuery data. It does not judge whether an agent's answers or decisions are correct.**

For security teams this closes a real gap: an agent runtime is a container running code with a cloud identity, and attackers who reach it through prompt injection or a poisoned dependency behave like attackers anywhere else. Here is what the service covers and where it stops.

## What are the basics?

| Item | Detail |
|---|---|
| Launch stage | Preview, under Google's Pre-GA Offerings Terms |
| Tiers | Security Command Center Premium; Enterprise tier is deprecated and shuts down on 21 May 2027 |
| What it watches | Supported agents on Agent Engine's runtime, plus related audit and runtime logs |
| Output | Findings in Security Command Center in near real time |
| Defaults | Runtime and control-plane detectors are enabled by default |

## Which runtime threats does it detect?

Runtime detectors look inside the agent's container. Examples from Google's detector list:

- **Execution:** added or modified malicious binaries, libraries or skills; malicious Python executed; container escape; fileless execution in `/memfd:`; Kubernetes attack tool execution; netcat remote code execution.
- **Reverse shells and scripts:** reverse shell, socat reverse shell, unexpected child shell, malicious script executed, malicious URL observed.
- **Credential access:** searching for Google Cloud credentials, private keys or passwords; GPG key reconnaissance.
- **Defense evasion:** Base64-encoded ELF, Python or shell scripts executed; launching a code compiler in the container.
- **Impact and exfiltration:** bulk data removal from disk, cryptocurrency mining over Stratum, launching remote file copy tools.
- **Privilege escalation:** known sudo and Polkit CVEs.

The runtime watcher can take up to a minute to start, and Google states data is processed in memory and stored only when reported as a finding.

## Which control-plane threats does it detect?

These detectors run through Event Threat Detection over audit logs and the agent runtime's output logs:

1. **Discovery:** port scanning from an AI agent; an agent's service account making unauthorized API calls or investigating its own permissions.
2. **Credential access:** anomalous agent access to the metadata service.
3. **Exfiltration:** agent-initiated BigQuery VPC Service Controls perimeter violations; BigQuery exfiltration to external tables or data extraction; Cloud SQL exports to public or external buckets.
4. **Initial access:** excessive permission-denied actions by an agent identity.
5. **Privilege escalation:** token generation using `signJwt`, implicit delegation, or cross-project OpenID and access tokens.

These map well to the agent identity model. See [Vertex AI Agent Engine agent identity](/blog/vertex-ai-agent-engine-agent-identity) for how agents get their own principals, and [VPC Service Controls for Vertex AI agents](/blog/vpc-service-controls-for-vertex-ai-agents) for the perimeter those exfiltration detectors watch.

## What does it not cover?

- **Agent behaviour.** Off-task answers, ungrounded claims, harmful outputs and policy violations in conversation are not infrastructure events.
- **Prompt injection itself.** It may catch the aftermath (a reverse shell, an odd token request), not the injected instruction.
- **Agents elsewhere.** Agents on other Google Cloud services, other clouds, SaaS platforms or developer laptops are out of scope.
- **Response.** Findings are alerts. Containing the agent is up to your playbook.

## How should you operationalise the findings?

1. Confirm your organization is on SCC Premium and plan for the Enterprise tier shutdown if relevant.
2. Keep detectors enabled; tune per module only after reviewing findings.
3. Export findings to your SIEM with the agent's resource name and service account attached. Our list of [AI agent events to send to a SIEM](/blog/ai-agent-events-to-send-to-siem) covers the other fields.
4. Map each finding type to a response: rotate credentials, block egress, quarantine or delete the agent.
5. Link findings to your agent inventory so the owner is paged, not just the cloud team.

For the wider control set, see [governing Vertex AI Agent Engine agents](/blog/govern-google-vertex-ai-agent-engine-agents) and the [platform guides pillar](/blog/how-to-govern-ai-agents-across-platforms).

## How Opencontroller does this

Opencontroller covers the behavioural layer this service does not. It auto-discovers agents across Google Cloud, AWS and Azure, has runtime adapters for Vertex AI, and applies live groundedness and task-adherence guardrails and LLM judges on production traffic. Rogue-agent handling monitors behavioural deviation, alerts, and can automatically quarantine and kill agents under estate-wide policies. [Book a demo](https://www.lyzr.ai/book-demo/).

## Frequently asked questions

**Is Agent Engine Threat Detection generally available?**
No. Google lists it as a preview offering under its Pre-GA terms. Check the Security Command Center docs for current status.

**Which Security Command Center tier do I need?**
Premium. The Enterprise tier also includes it but is deprecated and shuts down on 21 May 2027, with organizations moved to Premium.

**Does it detect prompt injection?**
Not directly. It detects infrastructure and identity activity that may follow a successful injection, such as reverse shells or unusual token generation.

## Key facts

- Google's Agent Engine Threat Detection (now documented as Agent Platform Threat Detection) is a Security Command Center preview service under Pre-GA terms, available in SCC Premium; the SCC Enterprise tier is deprecated and shuts down 21 May 2027.
- Runtime detectors (enabled by default) cover malicious binaries/libraries/skills, container escape, reverse shells, credential searches, Base64-encoded script execution, crypto mining and known sudo/Polkit CVEs; runtime watcher may take up to one minute to start.
- Control-plane detectors via Event Threat Detection cover agent port scanning, unauthorized service account API calls, anomalous metadata service access, BigQuery VPC-SC perimeter violations and exfiltration, Cloud SQL exports, excessive permission denials, and signJwt/implicit delegation/cross-project token generation.
- Findings appear in Security Command Center in near real time; data is stored only when reported as a finding.
- Opencontroller by Lyzr adds behavioural guardrails, LLM judges and rogue-agent quarantine across clouds; pricing starts from $250K per deployment per year.

## Related

- [How do you govern agents deployed on Vertex AI Agent Engine?](https://opencontroller.lyzr.com/blog/govern-google-vertex-ai-agent-engine-agents.md)
- [How does agent identity work in Vertex AI Agent Engine?](https://opencontroller.lyzr.com/blog/vertex-ai-agent-engine-agent-identity.md)
- [Which AI agent events should go to your SIEM?](https://opencontroller.lyzr.com/blog/ai-agent-events-to-send-to-siem.md)
- [More on Platform Guides](https://opencontroller.lyzr.com/blog/cluster/platform-guides)
