AI Runtime Post-Exploitation and Compute Theft
High severity TLP:CLEAR Part 2 of 2 Research by Rapid7 Published 19 Nov 2026
Hypothesis
An intruder has exploited an unauthenticated AI runtime to access system credentials via path traversal or hijacked cloud model subscriptions, evidenced by sensitive file access and unauthorized cloud API calls for premium models.
Brief
Why now
Rapid7 recently published a report titled Open by Default, Already Occupied: Threat Actors are Renting your Compute, which details how attackers target unauthenticated AI runtimes. These misconfigurations allow intruders to steal system credentials or run expensive inference tasks on the victim's infrastructure. While identifying open ports is a common task, identifying the specific transition from exposure to active abuse requires a deeper look at process and file behaviors.
How the hunt flows
The first phase scopes the estate using software inventory data. The query identifies hosts running known AI software like Ollama. This scoping step reduces the noise of subsequent checks by focusing the hunt on machines actually capable of hosting these workloads.
Once scoped, the hunt fans out into three parallel investigations across different surfaces. The first pivot examines process activity to stack-count command lines. It looks for rare model names or probe markers (such as "pwn" or "verify") that serve as visitor logs for attackers testing the runtime. The second pivot monitors file activity, specifically watching for the AI daemon accessing sensitive Linux paths like shadow files or SSH keys, which indicates a successful path-traversal exploit.
The third pivot shifts to cloud API activity. It identifies calls to high-value models like GPT-4 or Claude-3.1 originating from these workloads. This correlation detects LLMjacking, where an attacker uses a local AI gateway to consume a company's cloud credits. Finally, an agent triage step synthesizes these disparate signals to distinguish between a developer testing a local model and a malicious actor using the server as a proxy.
What the hunt cannot see
This hunt has two primary blind spots. Without syscall-level auditing, such as Linux auditd or eBPF-based file-read events, the hunt struggles to distinguish between a failed path-traversal attempt and a successful one. A process attempting to touch a sensitive file looks the same as a successful read in many telemetry snapshots. Second, if an attacker runs short-lived model inference processes between snapshot intervals, the hunt might miss the execution. Real-time event streams are required to guarantee capture of ephemeral compute theft.
In this series
Steps
-
Query, scoping
Scope AI runtime workloads
Identify hosts running Ollama or similar AI software to focus the subsequent behavioral checks.
sql, reads hb_software_inventorySELECT DISTINCT device_hostname, package_name, package_versionFROM hb_software_inventoryWHERE LOWER(package_name) LIKE '%ollama%'What a hit looks like. A list of hosts currently hosting the AI runtime. Silence suggests no managed installations, but unmanaged binaries might still exist.
-
Query, baseline
Rare Ollama command-line markers
Stack-count process command lines to find rare model pulls or names that act as visitor log markers.
sql, reads hb_process_activitySELECT LOWER(process_cmd_line) AS cmd, COUNT(DISTINCT device_hostname) AS hosts, COUNT(*) AS execs, MIN(time) AS first_seenFROM hb_process_activityWHERE (LOWER(process_name) LIKE '%ollama%' OR LOWER(process_cmd_line) LIKE '%ollama%') AND time >= datetime('now', '-{{lookback_days}} days') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0)GROUP BY cmdHAVING hosts <= 3ORDER BY hosts, execsWhat a hit looks like. Unique command lines containing tunnel domains (bore.pub, trycloudflare.com) or probe names (pwn, verify, beacon).
-
Query, detection candidate
AI runtime sensitive file access
Identify path-traversal payloads (CVE-2024-37032) used to read credentials or modify persistence locations.
sql, reads hb_file_activitySELECT device_hostname, process_name, file_path, activity_name, timeFROM hb_file_activityWHERE LOWER(process_name) LIKE '%ollama%' AND (instr(',' || '{{sensitive_files}}' || ',', ',' || LOWER(file_path) || ',') > 0 OR LOWER(file_path) LIKE '%/override.conf') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')What a hit looks like. The Ollama daemon touching /etc/shadow or its own Ed25519 identity key. This is a definitive sign of exploitation.
-
Agent triage
Analyze AI runtime compromise
Synthesize results from process, file, and cloud layers to confirm active exploitation or compute theft.
-
Decision
Route on workload verdict
Direct the response toward containment if the agent confirms malicious indicators.
-
Response action
Isolate compromised AI workload
Halt data exfiltration and compute theft by severing the host's network connectivity.
-
Response action
Revoke and rotate credentials
Invalidate potentially stolen credentials to prevent further access.
-
Analyst task
Analyst forensic review
Conduct a manual review of the evidence to determine the depth of the breach.
-
Analyst task
Harden AI runtimes
Finalize the hunt by ensuring residual AI workloads are secured.
Coverage
3 of 5 stages covered
| Stage and ATT&CK technique | Covered | How, or why not |
|---|---|---|
| Path Traversal Credential Access T1555 |
Yes | ollama-file-traversal, triage-ai-compromise |
| Execution Hijack and Persistence T1574.006 |
Yes | ollama-file-traversal, analyst-review |
| Resource Hijacking (LLMjacking) T1496 |
Yes | premium-model-usage, rare-ollama-commands, triage-ai-compromise |
| Unauthenticated AI API Exposure T1190 |
Out of scope | Belongs to another part of the 'Open by Default, Already Occupied: Threat Actors are Renting your Compute' series. |
| Outbound Callback and Tunneling T1572 |
Out of scope | Belongs to another part of the 'Open by Default, Already Occupied: Threat Actors are Renting your Compute' series. |
Blind spots
- Needs hb_linux_audit with file-read auditing. A failed exploit attempt looks identical to a successful one without detailed syscall auditing. It would answer: whether the sensitive file was successfully read or if the traversal attempt was blocked Remediation: Enable auditd or eBPF-based file-read auditing for the Ollama service account.
- Needs hb_process_activity with high-frequency event capture. Attackers may pull and run models in bursts between scanning intervals, which snapshots might miss. It would answer: whether short-lived model inference processes were missed by the snapshot tool Remediation: Transition to event-driven process monitoring to capture all execution events in real-time.
Parameters and data
Parameters
| Parameter | Type | Default | What it is |
|---|---|---|---|
lookback_days | number | 14 | Days of history to examine. |
premium_models | list[string] | gpt-4, gpt-4o, claude-3-opus, llama3.1:8b | High-value or signature model names used in compute-theft (LLMjacking) scenarios. |
scope_hosts | list[host] | none | List of hostnames to focus the hunt on; if empty, the hunt runs fleet-wide. |
sensitive_files | list[path] | /etc/shadow, /etc/passwd, /proc/self/environ, /etc/ld.so.preload, /.ssh/authorized_keys, /.ollama/id_ed25519 | Sensitive Linux files and Ollama keys often targeted for exfiltration or persistence. |
Data sources
| Source | Category | Telemetry |
|---|---|---|
| Endpoint telemetry (hb_ surfaces) | endpoint | endpoint |
| Identity / sign-in telemetry | identity | identity |
Source
An open hunt.md file. It runs anywhere that reads the format.
---
analysis: While a rule might alert on access to /etc/shadow, this hunt provides the
specific context of the AI runtime process, correlates it with the 'visitors' log'
found in rare command arguments, and assesses cloud financial impact which standard
rules cannot see.
blind_spots:
- id: no-linux-audit
owner: Endpoint Engineering
question: whether the sensitive file was successfully read or if the traversal attempt
was blocked
remediation: Enable auditd or eBPF-based file-read auditing for the Ollama service
account.
requires: hb_linux_audit with file-read auditing
risk: A failed exploit attempt looks identical to a successful one without detailed
syscall auditing.
stage: path-traversal-file-access
- id: ephemeral-compute
owner: SOC Architect
question: whether short-lived model inference processes were missed by the snapshot
tool
remediation: Transition to event-driven process monitoring to capture all execution
events in real-time.
requires: hb_process_activity with high-frequency event capture
risk: Attackers may pull and run models in bursts between scanning intervals, which
snapshots might miss.
stage: resource-hijacking-llmjacking
coverage:
- stage: path-traversal-file-access
status: covered
steps:
- ollama-file-traversal
- triage-ai-compromise
- reason: Monitors for writes to ld.so.preload and systemd override files used for
persistence.
stage: persistence-and-execution-hijack
status: covered
steps:
- ollama-file-traversal
- analyst-review
- stage: resource-hijacking-llmjacking
status: covered
steps:
- premium-model-usage
- rare-ollama-commands
- triage-ai-compromise
- reason: 'Belongs to another part of the ''Open by Default, Already Occupied: Threat
Actors are Renting your Compute'' series.'
stage: unauthenticated-api-exposure
status: out_of_scope
- reason: 'Belongs to another part of the ''Open by Default, Already Occupied: Threat
Actors are Renting your Compute'' series.'
stage: callback-verification-tunnels
status: out_of_scope
guardrails:
claims: no_unsupported
evidence: citation_required
missing_data: not_benign
telemetry: untrusted
hunt:
applicability: campaign-specific
handoff: keep-as-periodic-hunt
justification: Exposed AI runtimes provide a direct path to both compute resource
exhaustion (LLMjacking) and system credential theft. A negative result confirms
that these unmanaged workloads are not being used as open gateways for third-party
compute or data exfiltration.
methodology: model-assisted
trigger: intel-report
hypothesis: An intruder has exploited an unauthenticated AI runtime to access system
credentials via path traversal or hijacked cloud model subscriptions, evidenced
by sensitive file access and unauthorized cloud API calls for premium models.
labels:
- hunt
- attack.t1190
- attack.t1555
- attack.t1496
- attack.t1574.006
- attack.t1572
- command and control
- credential access
- impact
- initial access
- persistence
name: AI Runtime Post-Exploitation and Compute Theft
parameters:
lookback_days:
default: '14'
description: Days of history to examine.
from:
kind: manual
observed: '2024-05-20'
ref: standard-retention
type: number
premium_models:
default:
- gpt-4
- gpt-4o
- claude-3-opus
- llama3.1:8b
description: High-value or signature model names used in compute-theft (LLMjacking)
scenarios.
from:
kind: article
observed: '2026-11-18'
ref: https://www.rapid7.com/blog/post/tr-self-hosted-ai-renting-compute-exposed-servers
type: list[string]
scope_hosts:
default: []
description: List of hostnames to focus the hunt on; if empty, the hunt runs fleet-wide.
from:
kind: manual
observed: '2024-05-20'
ref: analyst-scoping
type: list[host]
sensitive_files:
default:
- /etc/shadow
- /etc/passwd
- /proc/self/environ
- /etc/ld.so.preload
- /.ssh/authorized_keys
- /.ollama/id_ed25519
description: Sensitive Linux files and Ollama keys often targeted for exfiltration
or persistence.
from:
kind: article
observed: '2026-11-18'
ref: https://www.rapid7.com/blog/post/tr-self-hosted-ai-renting-compute-exposed-servers
type: list[path]
provenance:
authors:
- name: Huntbase hunt generation
org: huntbase.io
generated:
by: huntbase-hunt-generation
from: https://www.rapid7.com/blog/post/tr-self-hosted-ai-renting-compute-exposed-servers
gates:
- dry-run
- lint
- critic
model: hb_google/gemini-3-flash-preview
rationale: Start with cloud-hosted Linux workloads in development or test accounts.
Widen to any server with public IP addresses and open port 11434.
references:
- name: "Rapid7 \u2014 Open by Default, Already Occupied: Threat Actors are Renting\
\ your Compute"
url: https://www.rapid7.com/blog/post/tr-self-hosted-ai-renting-compute-exposed-servers
related:
- hunt: unauthenticated-api-exposure-network-scoping
reason: This hunt focuses on post-exploitation behavior; identifying the open ports
belongs to a network exposure hunt.
relation: out-of-scope-alternative
- hunt: exposed-ai-runtime-probing-callbacks
relation: follows
scenario:
stages:
- name: Unauthenticated AI API Exposure
observables:
- port:11434
- port:8888
- port:8501
- port:8000
- url_path:/api/tags
- url_path:/api/pull
- url_path:/api/push
- url_path:/api/kernels
slug: unauthenticated-api-exposure
tactic: initial-access
techniques:
- T1190
- name: Outbound Callback and Tunneling
observables:
- domain:bore.pub
- domain:trycloudflare.com
- domain:ttl.sh
- domain:pinggy.io
- model_name:zzvitals-beacon
- model_name:verif_sys
- model_name:ops-verify
- model_name:*-copy-test
slug: callback-verification-tunnels
tactic: command-and-control
techniques:
- T1572
- name: Path Traversal Credential Access
observables:
- file_path:/etc/shadow
- file_path:/etc/passwd
- file_path:/proc/self/environ
- file_path:~/.ssh/authorized_keys
- file_path:~/.ollama/id_ed25519
- CVE-2024-37032
slug: path-traversal-file-access
tactic: credential-access
techniques:
- T1555
- name: Execution Hijack and Persistence
observables:
- file_path:/etc/ld.so.preload
- file_path:override.conf
- process_name:ollama
slug: persistence-and-execution-hijack
tactic: persistence
techniques:
- T1574.006
- name: Resource Hijacking (LLMjacking)
observables:
- model_name:gpt-4o
- model_name:claude-3-opus
- model_name:llama3.1:8b
slug: resource-hijacking-llmjacking
tactic: impact
techniques:
- T1496
summary: Threat actors are targeting unauthenticated Ollama AI servers to hijack
compute resources and linked cloud model subscriptions (LLMjacking). Beyond resource
theft, attackers use path traversal vulnerabilities (CVE-2024-37032) to read sensitive
host files like /etc/shadow and establish persistence via system configuration
hijacking.
series:
index: 2
slug: open-by-default-already-occupied-threat-actors-are-renting-your-compute
title: 'Open by Default, Already Occupied: Threat Actors are Renting your Compute'
total: 2
severity: high
targets:
analyst:
name: Tier-2 analyst
role: analyst
endpoint:
category: endpoint
name: Endpoint telemetry (hb_ surfaces)
telemetry:
- endpoint
hunter:
agent: true
name: Hunt agent
identity:
category: identity
name: Identity / sign-in telemetry
telemetry:
- identity
tlp: clear
type: investigation
---
# AI Runtime Post-Exploitation and Compute Theft
This hunt targets the post-exploitation phase of unauthenticated AI runtimes, specifically Ollama. It identifies attempts to read sensitive system files (CVE-2024-37032), such as shadow files or SSH keys, and detects the unauthorized usage of premium cloud models through hijacked AI gateway configurations. The hunt uses a funnel approach, scoping the estate for AI software before fanning out to detect rare command-line indicators, sensitive file access, and anomalous cloud API calls associated with high-value models. An agent then evaluates the combined evidence per host to differentiate between legitimate developer experimentation and malicious third-party activity.
## find-ai-workloads
<!-- Scope AI runtime workloads -->
Identify hosts running Ollama or similar AI software to focus the subsequent behavioral checks.
```sqlite target=endpoint role=scoping
~~~yaml
expected: A list of hosts currently hosting the AI runtime. Silence suggests no managed
installations, but unmanaged binaries might still exist.
reads:
- device_hostname
- package_name
- package_version
silence: not_evidence_of_absence
source: hb_software_inventory
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT DISTINCT device_hostname, package_name, package_version FROM hb_software_inventory WHERE LOWER(package_name) LIKE '%ollama%'
```
## parallel-investigation
<!-- Investigate activity side-by-side -->
parallel:
- → rare-ollama-commands
- → ollama-file-traversal
- → premium-model-usage
join: → triage-ai-compromise
## rare-ollama-commands
<!-- Rare Ollama command-line markers -->
Stack-count process command lines to find rare model pulls or names that act as visitor log markers.
```sqlite target=endpoint role=baseline params=(lookback_days=lookback_days, scope_hosts=scope_hosts)
~~~yaml
baseline:
compare: first_seen
window: '{{lookback_days}}d'
expected: Unique command lines containing tunnel domains (bore.pub, trycloudflare.com)
or probe names (pwn, verify, beacon).
prevalence:
by: device_hostname
key:
- process_cmd_line
rare_below: 3
reads:
- process_cmd_line
- device_hostname
- time
silence: not_evidence_of_absence
source: hb_process_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT LOWER(process_cmd_line) AS cmd, COUNT(DISTINCT device_hostname) AS hosts, COUNT(*) AS execs, MIN(time) AS first_seen FROM hb_process_activity WHERE (LOWER(process_name) LIKE '%ollama%' OR LOWER(process_cmd_line) LIKE '%ollama%') AND time >= datetime('now', '-{{lookback_days}} days') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) GROUP BY cmd HAVING hosts <= 3 ORDER BY hosts, execs
```
## ollama-file-traversal
<!-- AI runtime sensitive file access -->
Identify path-traversal payloads (CVE-2024-37032) used to read credentials or modify persistence locations.
```sqlite target=endpoint role=detection-candidate params=(sensitive_files=sensitive_files, scope_hosts=scope_hosts, lookback_days=lookback_days)
~~~yaml
expected: The Ollama daemon touching /etc/shadow or its own Ed25519 identity key.
This is a definitive sign of exploitation.
reads:
- device_hostname
- process_name
- file_path
- activity_name
- time
silence: not_evidence_of_absence
source: hb_file_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT device_hostname, process_name, file_path, activity_name, time FROM hb_file_activity WHERE LOWER(process_name) LIKE '%ollama%' AND (instr(',' || '{{sensitive_files}}' || ',', ',' || LOWER(file_path) || ',') > 0 OR LOWER(file_path) LIKE '%/override.conf') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')
```
## premium-model-usage
<!-- Anomalous premium model usage -->
Identify LLMjacking where unauthenticated runtimes are used to consume premium cloud model credits via the local AI gateway.
```sqlite target=endpoint role=enrichment params=(premium_models=premium_models, lookback_days=lookback_days)
~~~yaml
expected: Cloud API calls for premium models (e.g. gpt-4o, claude-3-opus) originating
from AI workloads that lack a legitimate business justification.
reads:
- cloud_account_uid
- actor_user_name
- resource_name
- api_operation
- time
silence: not_evidence_of_absence
source: hb_cloud_api_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT cloud_account_uid, actor_user_name, resource_name, api_operation, time FROM hb_cloud_api_activity WHERE instr(',' || '{{premium_models}}' || ',', ',' || LOWER(resource_name) || ',') > 0 AND time >= datetime('now', '-{{lookback_days}} days')
```
## triage-ai-compromise
<!-- Analyze AI runtime compromise -->
```agent target=hunter
cite: required
context:
- find-ai-workloads
- rare-ollama-commands
- ollama-file-traversal
- premium-model-usage
max_iterations: 6
objective: Determine if the identified AI workloads have been compromised by path-traversal
payloads or are being used for unauthorized cloud model inference by third parties.
success_criteria: Verdicts must distinguish between legitimate developer experimentation
and malicious third-party activity indicators.
tools:
- endpoint
```
## route-remediation
<!-- Route on workload verdict -->
if~: "the triage verdict is malicious for at least one host, indicating confirmed path traversal or LLMjacking" (confidence: high, judge=hunter)
then: → isolate-workload
indeterminate: → analyst-review
unavailable: → analyst-review (blind_spot: no-linux-audit)
else: → harden-runtimes
## isolate-workload
<!-- Isolate compromised AI workload -->
```action target=endpoint
~~~yaml
approval: required
~~~
Isolate the host and terminate the Ollama process immediately. Collect the .ollama directory and systemd unit logs for forensic review.
```
→ revoke-and-rotate
## revoke-and-rotate
<!-- Revoke and rotate credentials -->
```action target=identity
~~~yaml
approval: required
~~~
Revoke the Ollama Ed25519 identity key on ollama.com. Rotate any cloud access keys (AWS/OCI) discovered in the host's environment or config files.
```
→ analyst-review
## analyst-review
<!-- Analyst forensic review -->
```manual target=analyst
Review the cited rows and forensics data. Check SSH authorized_keys for unauthorized entries and examine the /etc/ld.so.preload file for persistence markers.
```
→ harden-runtimes
## harden-runtimes
<!-- Harden AI runtimes -->
```manual target=analyst
Ensure remaining Ollama instances are bound to localhost (OLLAMA_HOST=127.0.0.1) and implement egress controls to approved registries only.
```
→ end
Run this hunt against your own data.
In Huntbase, Scout runs every step across the tools you already connect, checks back through your history and returns a verdict with the evidence. The hunt.md file is open, so you can also take it anywhere that reads the format.
Drafted by machine (huntbase-hunt-generation, hb_google/gemini-3-flash-preview), checked by dry-run, lint, critic, then reviewed by a person.