review-security-k8s-agents-main · diff

git:20260521.e12360f to git:20260522.4bbd4a0

2 added, 2 removed. Audit A to A.

---

name: review-security-k8s-agents-main

description: Orchestrates comprehensive Kubernetes security audits specifically tailored for AI agent workloads.

---

# Task

Coordinate AI agent security review sub-agents, gather findings, and produce a summarized JSON report.

# Workflow

## 1. Context Ingestion

Pass project context (from `review-security-k8s-understand`) to sub-agents.

## 2. Parallel Reviews

- Launch in parallel:

+ Launch in parallel sub-agents:

- `review-security-k8s-agents-sandbox`

- `review-security-k8s-agents-firewall`

- `review-security-k8s-agents-credentials`

- `review-security-k8s-agents-prompt-injection`

- `review-security-k8s-agents-data-exfil`

- `review-security-k8s-agents-audit-logs`

**CRITICAL**: Instruct each to output JSON:

```json

[{"agent": "<skill-name>", "findings": [{"message": "<desc>", "file": "<name>", "line": "<num>"}]}]

```

(Return empty list if no findings). Wait for completion.

## 3. Triage & Filtering

Evaluate the raw findings against the project context to determine actual risk. Filter out findings that are functionally required by the workload's specific role or adequately mitigated by broader architectural controls.

- *Example:* Filter out missing egress proxy warnings if the agent's execution sandbox is completely air-gapped and the main control loop is strictly allowlisted to a single LLM API.

- - *Example:* Downgrade root execution warnings *inside* the execution sandbox if the context confirms the sandbox utilizes a secure VM-based `RuntimeClass` (e.g., Kata Containers) providing a hardware-level isolation boundary.

+ - *Example:* Filter out root execution warnings *inside* the execution sandbox if the context confirms the sandbox utilizes a secure VM-based `RuntimeClass` (e.g. gVisor or Kata Containers) providing a secure sandbox isolation boundary.

## 4. Aggregation

Merge the filtered findings into a single JSON array. Output MUST be valid JSON string (markdown blocks okay). Omit agents with no findings or return empty `findings`.