---
analysis: A single detection rule would struggle with the noise of legitimate software
  updates and shell usage. This hunt uses a risk-based lead to gate expensive queries,
  applies prevalence counting across the fleet to isolate rare drift, and uses an
  agent to correlate image risk with runtime behavior across three different telemetry
  surfaces.
blind_spots:
- id: limited-inventory-visibility
  question: Are there other unmanaged workloads outside of Databricks running these
    images?
  requires: Unified cloud inventory across AWS, Azure, and GCP
  risk: The scoping step is limited to Databricks; other container environments like
    EKS or GKE are not evaluated for the same risk in this hunt.
  stage: initial-access-exploit
- id: container-drift-blindness
  question: Are modifications occurring in temporary file layers that osquery cannot
    observe?
  requires: hb_file_activity with container layer awareness
  risk: If an adversary modifies files in a memory-backed overlay that is not flushed
    to the host disk, the file activity surface may not capture the drift.
  stage: persistence-through-drift
coverage:
- stage: initial-access-exploit
  status: covered
  steps:
  - identify-high-risk-images
  - evaluate-scoping-risk
- stage: runtime-container-execution
  status: covered
  steps:
  - suspicious-shell-launches
- stage: persistence-through-drift
  status: covered
  steps:
  - rare-binary-drift
- reason: "Belongs to another part of the 'Microsoft named a Leader in the Frost Radar\u2122\
    : Cloud Workload Protection Platforms, 2026' series."
  stage: over-permissioned-identity-access
  status: out_of_scope
- reason: "Belongs to another part of the 'Microsoft named a Leader in the Frost Radar\u2122\
    : Cloud Workload Protection Platforms, 2026' series."
  stage: outbound-command-and-control
  status: out_of_scope
guardrails:
  claims: no_unsupported
  evidence: citation_required
  missing_data: not_benign
  telemetry: untrusted
hunt:
  applicability: campaign-specific
  handoff: promote-to-detection
  justification: Modern cloud workloads are high-value targets where posture scanning
    is insufficient. Identifying exploitation at runtime through shell activity and
    binary drift is critical for stopping compromises that bypass initial deployment
    gates.
  methodology: model-assisted
  trigger: intel-report
hypothesis: An adversary has exploited a public-facing containerized application and
  is maintaining persistence through binary drift or suspicious shell execution within
  the workload runtime.
labels:
- hunt
- attack.t1190
- attack.t1059
- attack.t1542
name: Cloud Workload Runtime and Exploitation Behavior
parameters:
  lookback_days:
    default: '14'
    description: Days of history to examine for runtime behavior.
    from:
      kind: manual
      observed: '2026-08-19'
      ref: hunt-standard
    type: number
  scope_hosts:
    default: []
    description: Hostnames extracted from the scoping step to focus the runtime queries.
    from:
      kind: manual
      observed: '2026-08-19'
      ref: analyst-scoping
    type: list[host]
  vulnerable_tags:
    default:
    - latest
    - dev
    - test
    - old
    description: Image tags suspected of representing unversioned or high-risk development
      builds.
    from:
      kind: article
      observed: '2026-08-19'
      ref: frost-radar-2026
    type: list[string]
provenance:
  authors:
  - name: Huntbase hunt generation
    org: huntbase.io
  generated:
    by: huntbase-hunt-generation
    from: https://www.microsoft.com/en-us/security/blog/2026/08/19/microsoft-named-a-leader-in-the-frost-radar-cloud-workload-protection-platforms-2026/
    gates:
    - dry-run
    - lint
    model: hb_google/gemini-3-flash-preview
rationale: Start with Databricks compute clusters running containers with non-specific
  tags. These clusters represent the highest risk for unpatched, public-facing vulnerabilities.
references:
- name: 'Microsoft named a Leader in the Frost Radar: Cloud Workload Protection Platforms,
    2026'
  url: https://www.microsoft.com/en-us/security/blog/2026/08/19/microsoft-named-a-leader-in-the-frost-radar-cloud-workload-protection-platforms-2026/
related:
- hunt: over-permissioned-identity-access
  reason: That hunt focuses on the misuse of cloud identities after the workload has
    already been compromised.
  relation: out-of-scope-alternative
scenario:
  stages:
  - name: Exploit Public-Facing Application
    observables:
    - Incoming HTTP requests to vulnerable web applications
    - Presence of known vulnerabilities on internet-exposed assets
    - Misconfigured cloud infrastructure services
    slug: initial-access-exploit
    tactic: initial-access
    techniques:
    - T1190
  - name: Malicious Process Execution in Containers
    observables:
    - Suspicious process launches in Kubernetes pods
    - eBPF-detected anomalous runtime events
    - Interactive shell execution within running containers
    slug: runtime-container-execution
    tactic: execution
    techniques:
    - T1059
  - name: Container Binary Drift
    observables:
    - Unauthorized binary changes mid-run (drift)
    - Modifications to files within running container layers
    - Unexpected process activity from modified binaries
    slug: persistence-through-drift
    tactic: defense-evasion
    techniques:
    - T1542
  - name: Abuse of Over-Permissioned Identities
    observables:
    - Authentication using high-privilege service accounts
    - Anomalous sign-ins to cloud control planes (Azure, AWS, GCP)
    - Identity-linked access to Kubernetes API and resources
    slug: over-permissioned-identity-access
    tactic: credential-access
    techniques:
    - T1078
  - name: Outbound Command and Control
    observables:
    - DNS queries from Kubernetes pods to external domains
    - Outbound network connections to suspicious IP addresses
    - High-volume data transfer from container workloads
    slug: outbound-command-and-control
    tactic: command-and-control
    techniques:
    - T1071
  summary: This campaign involves the exploitation of public-facing applications to
    gain initial access to cloud workloads, particularly Kubernetes environments.
    Once inside, attackers execute malicious processes, leverage over-permissioned
    identities, and establish persistence through binary drift and outbound command-and-control
    traffic.
series:
  index: 1
  slug: microsoft-named-a-leader-in-the-frost-radar-cloud-workload-protection-platforms-2026
  title: "Microsoft named a Leader in the Frost Radar\u2122: Cloud Workload Protection\
    \ Platforms, 2026"
  total: 2
severity: medium
targets:
  analyst:
    name: Tier-2 analyst
    role: analyst
  databricks:
    category: siem
    huntbase:
      product: databricks
    name: databricks
  endpoint:
    category: endpoint
    name: Endpoint telemetry (hb_ surfaces)
    telemetry:
    - endpoint
  hunter:
    agent: true
    name: Hunt agent
tlp: clear
type: investigation
---


# Cloud Workload Runtime and Exploitation Behavior

This hunt targets the exploitation of public-facing cloud workloads by correlating high-risk container images with suspicious runtime behavior. It starts with a scoping query of Databricks clusters running images with generic or development tags, then gates deeper analysis on identified risks. The hunt subsequently fans out to find interactive shells spawned by web processes and rare file modifications within the container file system, providing a unified view of exploitation and drift.

## identify-high-risk-images
<!-- Identify high-risk container images -->
Find clusters running images with tags that bypass version control, representing a high risk for unpatched vulnerabilities.

```sqlite target=databricks role=scoping params=(vulnerable_tags=vulnerable_tags)
~~~yaml
expected: A list of clusters using non-versioned image tags. Silence suggests all
  production workloads use pinned, versioned tags.
reads:
- cluster_name
- docker_image
- creator_user_name
- cluster_id
silence: not_evidence_of_absence
source: databricks_compute_cluster
verified: dry-run
verified_at: '2026-09-20'
~~~
SELECT cluster_name, docker_image, creator_user_name, cluster_id FROM databricks_compute_cluster WHERE docker_image IS NOT NULL AND docker_image LIKE '%:%' AND instr(',' || '{{vulnerable_tags}}' || ',', ',' || REPLACE(LOWER(docker_image), SUBSTR(LOWER(docker_image), 1, INSTR(LOWER(docker_image), ':')), '') || ',') > 0
```

## evaluate-scoping-risk
<!-- Evaluate Scoping Risk -->
```agent target=hunter
cite: required
context:
- identify-high-risk-images
max_iterations: 3
objective: Review the images identified and determine if they represent production
  workloads. If high-risk clusters exist, provide the mapped device hostnames to be
  used for the scope_hosts parameter in the next phase.
success_criteria: A verdict naming high-risk hostnames to be scoped.
tools:
- databricks
- endpoint
```

## lead-gate
<!-- Lead Gate -->
if~: "The evaluate-scoping-risk verdict identifies at least one high-risk cluster hostname." (confidence: high, judge=hunter)
then: → runtime-fan-out
indeterminate: → close-out-negative
unavailable: → close-out-negative (blind_spot: limited-inventory-visibility)
else: → close-out-negative

## runtime-fan-out
<!-- Runtime Evidence Gathering -->
parallel:
- → suspicious-shell-launches
- → rare-binary-drift
join: → runtime-triage

## suspicious-shell-launches
<!-- Suspicious Shell Launches -->
Detect interactive shells spawned by web applications, a primary indicator of successful RCE.

```sqlite target=endpoint role=detection-candidate params=(scope_hosts=scope_hosts, lookback_days=lookback_days)
~~~yaml
expected: A shell process whose parent is a web-facing service. Silence confirms no
  common RCE patterns are active.
reads:
- device_hostname
- process_name
- process_cmd_line
- parent_process_name
- time
- on_disk
silence: evidence_of_absence
source: hb_process_activity
verified: dry-run
verified_at: '2026-09-20'
~~~
SELECT device_hostname, process_name, process_cmd_line, parent_process_name, time FROM hb_process_activity WHERE ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND (on_disk = 0 OR (LOWER(process_name) LIKE '%sh' AND (LOWER(parent_process_name) LIKE '%java%' OR LOWER(parent_process_name) LIKE '%python%' OR LOWER(parent_process_name) LIKE '%node%' OR LOWER(parent_process_name) LIKE '%php%'))) AND time >= datetime('now', '-{{lookback_days}} days')
```

## rare-binary-drift
<!-- Rare Binary Drift in Containers -->
Identify unauthorized changes to system binaries on a small number of hosts, suggesting persistence.

```sqlite target=endpoint role=baseline params=(scope_hosts=scope_hosts, lookback_days=lookback_days)
~~~yaml
baseline:
  compare: first_seen
  window: '{{lookback_days}}d'
expected: File writes to protected system paths occurring on only one or two hosts.
  Fleet-wide writes are likely updates and ignored.
prevalence:
  by: device_hostname
  key:
  - file_path
  rare_below: 3
reads:
- file_path
- device_hostname
- activity_id
- time
silence: not_evidence_of_absence
source: hb_file_activity
verified: dry-run
verified_at: '2026-09-20'
~~~
SELECT LOWER(file_path) AS path, COUNT(DISTINCT device_hostname) AS host_count, MIN(time) AS first_seen FROM hb_file_activity WHERE ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND activity_id IN (1, 3, 5) AND (LOWER(file_path) LIKE '/bin/%' OR LOWER(file_path) LIKE '/usr/bin/%' OR LOWER(file_path) LIKE '/etc/%') AND time >= datetime('now', '-{{lookback_days}} days') GROUP BY path HAVING host_count <= 2 ORDER BY host_count ASC
```

## runtime-triage
<!-- Runtime Triage -->
```agent target=hunter
cite: required
context:
- evaluate-scoping-risk
- suspicious-shell-launches
- rare-binary-drift
max_iterations: 5
objective: Determine if the combination of a high-risk container image, a suspicious
  shell launch, and rare binary drift indicates a successful compromise. Citing specific
  hosts and processes is required.
success_criteria: A verdict of malicious | suspicious | benign per host.
tools:
- databricks
- endpoint
```

## final-routing
<!-- Final Routing -->
if~: "The runtime-triage verdict is malicious for at least one host." (confidence: high, judge=hunter)
then: → isolate-workload
indeterminate: → analyst-final-review
unavailable: → analyst-final-review (blind_spot: container-drift-blindness)
else: → analyst-final-review

## isolate-workload
<!-- Isolate Workload -->
```action target=endpoint
~~~yaml
approval: required
~~~
Isolate the compromised host node using the EDR control plane. Do not terminate running pods until a memory dump of the suspicious process is captured.
```
→ analyst-final-review

## analyst-final-review
<!-- Analyst Final Review -->
```manual target=analyst
Examine the rare file writes and shell execution context. Verify if the activity corresponds to known maintenance. Coordinate with the cluster owner to redeploy using a hardened, versioned Docker image.
```
→ close-out-hunt

## close-out-negative
<!-- Close Out (No Findings) -->
```manual target=analyst
Record that the current cluster inventory uses versioned images or that no high-risk tags were found. File a recommendation to enforce image pinning at the policy level.
```
→ end

## close-out-hunt
<!-- Close Out Hunt -->
```manual target=analyst
Update the security policy based on findings. If binary drift was confirmed, work with the platform team to enable immutable infrastructure controls.
```
→ end