Exposed AI Runtime Probing and Callbacks
High severity TLP:CLEAR Part 1 of 2 Research by Rapid7 Published 19 Nov 2026
Hypothesis
An adversary is identifying unauthenticated AI runtimes exposed to the internet and using them to verify connectivity or establish command-and-control via outbound tunneling services.
Brief
Why Now
Rapid7 recently published Open by Default, Already Occupied: Threat Actors are Renting your Compute, detailing how adversaries exploit unauthenticated AI runtimes. Thousands of instances of Ollama and Jupyter are currently exposed to the internet without a password. Attackers do not just want the data; they want the compute. They use these systems to run their own models or establish persistent outbound tunnels to bypass firewalls.
How the Hunt Flows
The hunt begins with a scoping phase on the hb_exposed_assets surface. It identifies internal hosts that external scanners have already flagged as exposing AI runtimes. This focuses the investigation on the most likely entry points rather than searching every host in the estate.
Once the hunt identifies exposed hosts, it pivots to behavioral telemetry. The workflow runs two queries in parallel. The first checks the hb_dns_activity surface for resolutions of free tunneling domains like bore.pub or trycloudflare.com. These services are commonly used by attackers to create a stable callback from an unauthenticated server. The second query monitors the hb_http_activity surface for probes against sensitive API paths, such as /api/pull or /api/tags, coming from external IP addresses.
Finally, an analyst or agent triage weighs these signals. A host that appears in the external exposure list and also initiates outbound tunnel DNS lookups suggests a successful compromise. This correlation allows the team to distinguish between a service that is merely misconfigured and one that is actively being used as an adversary's remote terminal.
What the Hunt Cannot See
This hunt has two primary blind spots. First, it cannot inspect the full HTTP request body in many standard proxy logs. An attacker might use a POST request to /api/pull containing a path-traversal payload; while we see the request to the endpoint, we may not see the specific malicious manifest digest. Second, unmanaged VPS hosts often lack process-to-network mapping. Without this, we cannot definitively prove the Ollama process initiated the DNS resolution rather than a legitimate developer tool.
In this series
Steps
-
Query, scoping
Identify exposed AI services
Find hosts that external scanners have already flagged as exposing AI runtimes or related ports.
sql, reads hb_exposed_assetsSELECT domain_or_ip, port, product, providerFROM hb_exposed_assetsWHERE (LOWER(product) LIKE '%ollama%' OR LOWER(product) LIKE '%jupyter%' OR port IN (11434, 8888, 8501, 8000))What a hit looks like. Rows naming external IPs and ports of known AI runtimes. Silence suggests no current exposure detected by external scanning.
-
Query, baseline
DNS resolution for tunnel services
Detect internal hosts resolving tunnel providers which indicates a successful callback or verification check.
sql, reads hb_dns_activitySELECT device_hostname, query_hostname, COUNT(*) AS lookups, MIN(time) AS first_seenFROM hb_dns_activityWHERE instr(',' || '{{tunnel_domains}}' || ',', ',' || LOWER(query_hostname) || ',') > 0 AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')GROUP BY device_hostname, query_hostnameWhat a hit looks like. Internal hosts attempting to connect to free tunnel providers. Silence means no known tunnel callbacks were initiated.
-
Query, detection candidate
Unauthenticated AI API probing
Identify external requests to sensitive endpoints that expose server state or model information across several runtimes.
sql, reads hb_http_activitySELECT device_hostname, url_path, src_endpoint_ip, status_code, timeFROM hb_http_activityWHERE (LOWER(url_path) LIKE '/api/tags%' OR LOWER(url_path) LIKE '/api/pull%' OR LOWER(url_path) LIKE '/api/push%' OR LOWER(url_path) LIKE '/api/kernels%' OR LOWER(url_path) LIKE '/api/contents%' OR LOWER(url_path) LIKE '/api/sessions%') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')What a hit looks like. HTTP requests from external IPs to administrative paths. Silence suggests no unauthenticated probing was logged.
-
Agent triage
Weigh exposure and callbacks
Determine if the unauthenticated exposure and subsequent tunnel activity indicate a successful exploitation.
-
Decision
Route on verdict
Direct the hunt based on the agent's assessment of compromise.
-
Response action
Isolate compromised AI host
Stop further misuse or data exfiltration from the compromised runtime.
-
Analyst task
Manual forensic review
Identify specific model names or encoded payloads that indicate the attacker's intent.
-
Analyst task
Remediation and hardening
Close the exposure gap for hosts that were not actively misused.
Coverage
2 of 5 stages covered
| Stage and ATT&CK technique | Covered | How, or why not |
|---|---|---|
| Unauthenticated AI API Exposure T1190 |
Yes | exposed-ai-scoping, suspicious-api-probes |
| Outbound Callback and Tunneling T1572 |
Yes | outbound-tunnel-dns |
| Path Traversal Credential Access T1555 |
Out of scope | This belongs to the second hunt in the series focusing on post-exploitation file access. |
| Execution Hijack and Persistence T1574.006 |
Out of scope | Belongs to a separate hunt focusing on ld.so.preload and systemd persistence. |
| Resource Hijacking (LLMjacking) T1496 |
Out of scope | Belongs to a hunt focusing on cloud billing and subscription abuse. |
Blind spots
- Needs hb_http_activity with full request body capture. A simple GET to /api/tags is visible, but the destructive manifest digest payload in a POST /api/pull body is often missed by standard proxy logs. It would answer: whether an /api/pull request contained a path-traversal payload Remediation: Enable verbose logging for reverse proxies fronting AI infrastructure.
- Needs hb_network_connection with process mapping on all VPS hosts. Without process-to-network mapping on unmanaged VPS hosts, we cannot confirm the AI runtime initiated the callback versus a standard developer tool. It would answer: which specific process initiated the connection to a tunnel provider Remediation: Deploy endpoint agents to all expensed cloud compute.
Parameters and data
Parameters
| Parameter | Type | Default | What it is |
|---|---|---|---|
lookback_days | number | 14 | Days of history to examine for behavioural telemetry. |
scope_hosts | list[host] | none | Optional list of hosts to narrow the behavioural search; leave empty to hunt across the whole estate. |
tunnel_domains | list[domain] | bore.pub, trycloudflare.com, ttl.sh, pinggy.io | Free tunneling and throwaway registry services used for callbacks and reachability verification. |
Data sources
| Source | Category | Telemetry |
|---|---|---|
| Endpoint telemetry (hb_ surfaces) | endpoint | endpoint |
| Web server / proxy logs | siem | network |
Source
An open hunt.md file. It runs anywhere that reads the format.
---
analysis: A single rule on Ollama's port cannot distinguish between a researcher's
honeypot and a corporate asset. This hunt correlates external scan data with internal
behavioural pivots (DNS and API probing) to settle on a compromise verdict.
blind_spots:
- id: missing-http-request-bodies
owner: Infrastructure Team
question: whether an /api/pull request contained a path-traversal payload
remediation: Enable verbose logging for reverse proxies fronting AI infrastructure.
requires: hb_http_activity with full request body capture
risk: A simple GET to /api/tags is visible, but the destructive manifest digest
payload in a POST /api/pull body is often missed by standard proxy logs.
stage: unauthenticated-api-exposure
- id: egress-telemetry-gap
owner: Detection Engineering
question: which specific process initiated the connection to a tunnel provider
remediation: Deploy endpoint agents to all expensed cloud compute.
requires: hb_network_connection with process mapping on all VPS hosts
risk: Without process-to-network mapping on unmanaged VPS hosts, we cannot confirm
the AI runtime initiated the callback versus a standard developer tool.
stage: callback-verification-tunnels
coverage:
- stage: unauthenticated-api-exposure
status: covered
steps:
- exposed-ai-scoping
- suspicious-api-probes
- stage: callback-verification-tunnels
status: covered
steps:
- outbound-tunnel-dns
- reason: This belongs to the second hunt in the series focusing on post-exploitation
file access.
stage: path-traversal-file-access
status: out_of_scope
- reason: Belongs to a separate hunt focusing on ld.so.preload and systemd persistence.
stage: persistence-and-execution-hijack
status: out_of_scope
- reason: Belongs to a hunt focusing on cloud billing and subscription abuse.
stage: resource-hijacking-llmjacking
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: Thousands of AI runtimes are exposed unauthenticated; the model list
itself becomes a visitors' log of attacker activity. A negative result confirms
that your compute is not being 'rented' by strangers.
methodology: model-assisted
trigger: intel-report
hypothesis: An adversary is identifying unauthenticated AI runtimes exposed to the
internet and using them to verify connectivity or establish command-and-control
via outbound tunneling services.
labels:
- hunt
- attack.t1190
- attack.t1572
- command and control
- credential access
- impact
- initial access
- persistence
name: Exposed AI Runtime Probing and Callbacks
parameters:
lookback_days:
default: '14'
description: Days of history to examine for behavioural telemetry.
from:
kind: manual
observed: '2024-11-18'
ref: standard-lookback
type: number
scope_hosts:
default: []
description: Optional list of hosts to narrow the behavioural search; leave empty
to hunt across the whole estate.
from:
kind: manual
observed: '2024-11-18'
ref: analyst-scoping
type: list[host]
tunnel_domains:
default:
- bore.pub
- trycloudflare.com
- ttl.sh
- pinggy.io
description: Free tunneling and throwaway registry services used for callbacks
and reachability verification.
from:
kind: article
observed: '2026-11-18'
ref: rapid7-exposed-ai
type: list[domain]
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
model: hb_google/gemini-3-flash-preview
rationale: Focus on cloud and hosting network ranges (OVH, Hetzner, DigitalOcean)
where developers may have 'expensed' servers outside of standard CMDB oversight.
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: ollama-path-traversal-manifest-digest
reason: This hunt focuses on the exposure and network plane; path traversal exploitation
is a separate behavioral hunt on the host surface.
relation: out-of-scope-alternative
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: 1
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
web:
category: siem
name: Web server / proxy logs
telemetry:
- network
tlp: clear
type: investigation
---
# Exposed AI Runtime Probing and Callbacks
This hunt identifies AI runtimes like Ollama and Jupyter exposed to the public internet without authentication. It then searches for evidence of active misuse by correlating this exposure with outbound DNS callbacks to free tunneling services and suspicious HTTP probes against sensitive API endpoints. By combining external exposure data with internal telemetry, the hunt distinguishes between simple misconfiguration and active server-side request forgery (SSRF) or persistence attempts.
## exposed-ai-scoping
<!-- Identify exposed AI services -->
Find hosts that external scanners have already flagged as exposing AI runtimes or related ports.
```sqlite target=endpoint role=scoping
~~~yaml
expected: Rows naming external IPs and ports of known AI runtimes. Silence suggests
no current exposure detected by external scanning.
reads:
- domain_or_ip
- port
- product
- provider
silence: not_evidence_of_absence
source: hb_exposed_assets
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT domain_or_ip, port, product, provider FROM hb_exposed_assets WHERE (LOWER(product) LIKE '%ollama%' OR LOWER(product) LIKE '%jupyter%' OR port IN (11434, 8888, 8501, 8000))
```
## corroborate-misuse
<!-- Corroborate misuse side by side -->
parallel:
- → outbound-tunnel-dns
- → suspicious-api-probes
join: → triage-ai-exposure
## outbound-tunnel-dns
<!-- DNS resolution for tunnel services -->
Detect internal hosts resolving tunnel providers which indicates a successful callback or verification check.
```sqlite target=endpoint role=baseline params=(tunnel_domains=tunnel_domains, scope_hosts=scope_hosts, lookback_days=lookback_days)
~~~yaml
baseline:
compare: first_seen
window: '{{lookback_days}}d'
expected: Internal hosts attempting to connect to free tunnel providers. Silence means
no known tunnel callbacks were initiated.
prevalence:
by: device_hostname
key:
- query_hostname
rare_below: 3
reads:
- device_hostname
- query_hostname
- time
silence: not_evidence_of_absence
source: hb_dns_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT device_hostname, query_hostname, COUNT(*) AS lookups, MIN(time) AS first_seen FROM hb_dns_activity WHERE instr(',' || '{{tunnel_domains}}' || ',', ',' || LOWER(query_hostname) || ',') > 0 AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days') GROUP BY device_hostname, query_hostname
```
## suspicious-api-probes
<!-- Unauthenticated AI API probing -->
Identify external requests to sensitive endpoints that expose server state or model information across several runtimes.
```sqlite target=web role=detection-candidate params=(scope_hosts=scope_hosts, lookback_days=lookback_days)
~~~yaml
expected: HTTP requests from external IPs to administrative paths. Silence suggests
no unauthenticated probing was logged.
reads:
- device_hostname
- url_path
- src_endpoint_ip
- status_code
- time
silence: not_evidence_of_absence
source: hb_http_activity
verified: dry-run
verified_at: '2026-11-19'
~~~
SELECT device_hostname, url_path, src_endpoint_ip, status_code, time FROM hb_http_activity WHERE (LOWER(url_path) LIKE '/api/tags%' OR LOWER(url_path) LIKE '/api/pull%' OR LOWER(url_path) LIKE '/api/push%' OR LOWER(url_path) LIKE '/api/kernels%' OR LOWER(url_path) LIKE '/api/contents%' OR LOWER(url_path) LIKE '/api/sessions%') AND ('{{scope_hosts}}' = '' OR instr(',' || '{{scope_hosts}}' || ',', ',' || device_hostname || ',') > 0) AND time >= datetime('now', '-{{lookback_days}} days')
```
## triage-ai-exposure
<!-- Weigh exposure and callbacks -->
```agent target=hunter
cite: required
context:
- exposed-ai-scoping
- outbound-tunnel-dns
- suspicious-api-probes
max_iterations: 4
objective: Determine if any host with an exposed AI service has also resolved tunnel
domains or received suspicious API requests indicating misuse.
success_criteria: A verdict of malicious | suspicious | benign for each host, citing
the external probe and the tunnel resolution.
tools:
- endpoint
- web
```
## decide-compromise
<!-- Route on verdict -->
if~: "the triage-ai-exposure verdict is malicious for at least one host" (confidence: high, judge=hunter)
then: → isolate-host
indeterminate: → manual-forensics
unavailable: → manual-forensics (blind_spot: missing-http-request-bodies)
else: → remediation-hardening
## isolate-host
<!-- Isolate compromised AI host -->
```action target=endpoint
~~~yaml
approval: required
~~~
Isolate the host immediately. Revoke any linked Ollama account keys and check for unauthorized cloud spend.
```
→ manual-forensics
## manual-forensics
<!-- Manual forensic review -->
```manual target=analyst
Read the model list on the affected host using the Ollama or Jupyter API. Look for model names prefixed with IP addresses or containing Base64 strings that decode to system paths.
```
→ end
## remediation-hardening
<!-- Remediation and hardening -->
```manual target=analyst
Bind the AI runtime to localhost or a private network. Implement an authenticating reverse proxy or zero-trust layer for remote access.
```
→ 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, then reviewed by a person.