Industrial cybersecurity firm Dragos disclosed this week that attackers maintained persistent access inside three separate AI inference providers for months, exfiltrating customer prompt data — and none of the three providers detected the intrusion on their own. They only learned of the compromise when researchers notified them. OpenRouter, the popular AI gateway that routes customer requests to a pool of backend inference providers, has dropped all three from its platform.
The detail that should stop every CISO cold: some of the stolen requests were flagged as Zero Data Retention (ZDR) — the contractual setting that is supposed to guarantee a prompt is never written to disk, never logged, never retained. If ZDR-flagged prompts were accessible to an intruder, they were being stored somewhere they should not have been, or they were being intercepted in transit.
If your organization sends any prompts through OpenRouter, an inference API, or any LLM gateway — code, customer data, legal documents, source material, internal strategy — you need to assume downstream provider compromise is a realistic threat scenario, not a theoretical one. This post breaks down what happened, how to audit your own exposure, and how to build detections for the exact behaviors that went unnoticed for months.
Technical Analysis
What happened
Per the Dragos disclosure reported by The Hacker News:
- Attackers gained and maintained access to the internal environments of three AI inference providers over a period of months.
- Customer inference requests (prompts) were stolen from inside these environments.
- A subset of the stolen requests carried Zero Data Retention flags, meaning customers had explicitly configured their requests to never be persisted.
- The providers had no internal detection of the intrusion; discovery came from external researchers.
- OpenRouter, upon notification, removed all three providers from its routing pool.
Why this attack pattern matters for defenders
This is not a classic perimeter exploit story — it is a third-party / supply-chain compromise of the AI inference layer, and it exposes a systemic blind spot:
- The inference provider is an implicit data processor. Every prompt you send — including API keys pasted by developers, PII, source code, and customer records — transits and is processed by infrastructure you do not control and cannot monitor directly.
- ZDR is a policy, not a control. A ZDR flag is only meaningful if the provider's infrastructure actually enforces it: no request logging at the proxy tier, no debug logging in the inference stack, no analytics pipelines capturing request bodies, no ephemeral storage that outlives the request. The theft of ZDR-flagged prompts indicates the data existed somewhere in a retrievable form.
- Detection lag is the real story. Months-long dwell time at three providers simultaneously suggests attackers are deliberately targeting AI infrastructure as a high-value intelligence collection source — and that most providers lack mature internal threat detection for their own inference and logging pipelines.
The attack chain (defender's view)
While Dragos has not published full TTPs at the time of writing, the observable pattern for long-dwell data theft at an inference provider typically follows this chain:
- Initial access — exposed management interfaces, compromised credentials, or a vulnerable service on the provider's infrastructure.
- Persistence and access to the data plane — the attacker reaches systems where request/response traffic is processed or logged: API gateways, reverse proxies, model-serving hosts, telemetry/observability pipelines, or object storage backing analytics.
- Collection — bulk access to prompt/response logs, or passive capture of inference traffic. The ZDR violation implies collection from a store that should not exist or interception before any retention logic is applied.
- Exfiltration — low-and-slow egress to attacker infrastructure, often disguised within legitimate cloud traffic volumes (inference providers push enormous legitimate traffic, which is exactly why the theft blended in).
Exploitation status
- Confirmed active intrusion and data theft at three providers, per Dragos.
- No CVE has been publicly attributed to this campaign; there is no single patch to apply. The defensive problem is architectural: visibility, egress control, and verification of data-handling claims at third parties.
Detection & Response
The detections below serve two audiences: (1) LLM/inference operators and platform teams who need to detect the exact failure modes that went unnoticed — unauthorized access to prompt stores, ZDR policy violations, and anomalous egress; and (2) enterprise defenders who need to hunt their own AI gateway traffic for signs of downstream exposure.
Audit First: The ZDR Integrity Check
The single most important immediate action if you operate inference infrastructure: prove that ZDR-flagged requests are not persisted anywhere. Search your log pipelines, object storage, and observability platforms for request bodies that carry ZDR metadata. If you find them, you have a policy violation — and possibly the same exposure Dragos' targets had.
Sigma Rules
The following rules target the behaviors that matter: unexpected processes reading inference log stores, and anomalous bulk reads of API/proxy access logs. Tune paths to your environment — the goal is scoping reads of prompt/log data to a known allowlist of service accounts and binaries.
---
title: Unauthorized Process Access to LLM Inference or Prompt Log Stores
id: 8f2c1a94-3b7d-4e5f-9a61-2c8d4e7b0a53
status: experimental
description: Detects non-standard processes reading directories or files associated with LLM inference request logs, proxy logs, or prompt stores. Long-dwell intrusions at AI providers involved quiet access to request data; reads of these paths should be restricted to known service binaries.
references:
- https://thehackernews.com/2026/11/openrouter-drops-3-providers-after.html
- https://attack.mitre.org/techniques/T1005/
author: Security Arsenal
date: 2026/11/06
tags:
- attack.collection
- attack.t1005
logsource:
category: file_event
product: linux
detection:
selection_paths:
TargetFilename|contains:
- '/var/log/nginx/'
- '/var/log/envoy/'
- '/var/log/openrouter/'
- '/var/log/llm/'
- '/var/log/inference/'
- '/var/log/vllm/'
- '/opt/llm/logs/'
filter_known_services:
Image|endswith:
- '/rsyslogd'
- '/fluent-bit'
- '/vector'
- '/filebeat'
- '/otelcol'
condition: selection_paths and not filter_known_services
falsepositives:
- Backup agents and log shippers not yet in the allowlist
- Legitimate debugging by SRE staff during incidents
level: high
---
title: Interactive Shell Spawning From AI Serving or Proxy Processes
id: 3d9e7b21-6a4f-4c8d-b5e2-7f1a9c3d8e64
status: experimental
description: Detects interactive shells spawned as children of model-serving, API gateway, or reverse proxy processes. Inference stacks (vLLM, Triton, TGI, nginx, Envoy) should never spawn shells; this is a hallmark of webshell-style or post-exploitation activity on the data plane.
references:
- https://thehackernews.com/2026/11/openrouter-drops-3-providers-after.html
- https://attack.mitre.org/techniques/T1059/
author: Security Arsenal
date: 2026/11/06
tags:
- attack.execution
- attack.t1059.004
logsource:
category: process_creation
product: linux
detection:
selection_parents:
ParentImage|endswith:
- '/vllm'
- '/tritonserver'
- '/text-generation-server'
- '/nginx'
- '/envoy'
- '/gunicorn'
- '/uvicorn'
selection_children:
Image|endswith:
- '/bash'
- '/sh'
- '/zsh'
- '/python'
- '/python3'
- '/curl'
- '/wget'
condition: selection_parents and selection_children
falsepositives:
- Health-check scripts invoked by misconfigured service wrappers
- Container entrypoint logic during startup (scope to post-startup windows if needed)
level: critical
KQL Hunt — Egress Anomaly and Prompt-Store Access (Microsoft Sentinel)
This query hunts for two patterns ingested via Syslog/CEF or Defender: (1) outbound transfers from inference/gateway hosts to rare external destinations — the low-and-slow exfiltration that hid inside legitimate traffic for months — and (2) processes touching inference log paths. Run it over a 30-day lookback; the dwell time in this campaign was measured in months.
// Hunt: anomalous egress from LLM gateway/inference hosts + access to prompt log stores
// Tables: DeviceNetworkEvents, DeviceProcessEvents (adjust host list to your AI gateway/serving fleet)
let AiHosts = dynamic(["llm-gw-01", "inference-node-01", "vllm-prod-01"]); // replace with your AI gateway/inference host inventory
let Lookback = 30d;
let BaselineDest =
DeviceNetworkEvents
| where TimeGenerated > ago(Lookback * 2) and TimeGenerated < ago(Lookback)
| where DeviceName in~ (AiHosts)
| where RemoteIPType == "Public"
| distinct RemoteIP, RemoteUrl;
DeviceNetworkEvents
| where TimeGenerated > ago(Lookback)
| where DeviceName in~ (AiHosts)
| where RemoteIPType == "Public"
| where RemoteIP !in (BaselineDest)
| summarize
FirstSeen = min(TimeGenerated),
LastSeen = max(TimeGenerated),
Connections = count(),
Processes = make_set(InitiatingProcessFileName),
Accounts = make_set(InitiatingProcessAccountName)
by DeviceName, RemoteIP, RemoteUrl, RemotePort
| where Connections >= 5
| sort by FirstSeen asc;
// Secondary hunt: processes reading inference/proxy log paths
DeviceProcessEvents
| where TimeGenerated > ago(Lookback)
| where DeviceName in~ (AiHosts)
| where ProcessCommandLine has_any ("/var/log/nginx", "/var/log/envoy", "/var/log/inference", "/var/log/vllm", "access.log", "prompt")
| where FileName !in~ ("rsyslogd", "fluent-bit", "filebeat", "vector", "otelcol")
| project TimeGenerated, DeviceName, FileName, ProcessCommandLine, AccountName, InitiatingProcessFileName
| sort by TimeGenerated desc;
Velociraptor VQL — Hunt Access to Prompt Stores and Outbound Connections
For DFIR teams validating whether an inference or gateway host has been touched, this artifact enumerates processes with open connections to external IPs alongside any process whose command line references log store paths — the two artifacts that would have surfaced this campaign.
-- Artifact: AI Inference Host Compromise Triage
-- Surfaces processes reading prompt/log stores and live outbound connections
-- from serving processes on LLM gateway / inference nodes.
LET log_readers = SELECT Pid, Name, CommandLine, Exe, Username, CreateTime
FROM pslist()
WHERE CommandLine =~ '(access\\.log|/var/log/nginx|/var/log/envoy|/var/log/inference|/var/log/vllm|prompt|completion)'
AND Name !~ '(rsyslogd|fluent-bit|filebeat|vector|otelcol)'
LET suspicious_conns = SELECT Pid, Name, Ppid, Status, RemoteAddress, RemotePort
FROM netstat()
WHERE Status =~ 'ESTABLISHED'
AND RemoteAddress !~ '^(10\\.|172\\.(1[6-9]|2[0-9]|3[01])\\.|192\\.168\\.|127\\.|::1)'
AND Name =~ '(vllm|triton|text-generation|nginx|envoy|python|uvicorn|gunicorn)'
SELECT 'log_store_reader' AS FindingType, Pid, Name, CommandLine AS Detail,
Exe AS Binary, Username, CreateTime AS Timestamp, '' AS RemoteAddress, 0 AS RemotePort
FROM log_readers
UNION ALL
SELECT 'outbound_connection' AS FindingType, Pid, Name, '' AS Detail,
'' AS Binary, '' AS Username, now() AS Timestamp, RemoteAddress, RemotePort
FROM suspicious_conns
Remediation / Hardening Script — ZDR Compliance Audit and Egress Baseline
For platform teams operating inference infrastructure, this Bash script audits whether request bodies are being persisted where they should not be, checks for unexpected listeners and outbound connections on serving hosts, and inventories log-rotation gaps that let sensitive data accumulate.
#!/bin/bash
# zdr_audit.sh — Verify Zero Data Retention enforcement and baseline egress on LLM serving hosts
# Run on each inference/gateway node. Review output; do not auto-remediate in production.
set -euo pipefail
REPORT="zdr_audit_$(hostname)_$(date +%Y%m%d_%H%M%S).txt"
{
echo "=== ZDR / Prompt Retention Audit — $(hostname) — $(date -u) ==="
echo "--- [1] Searching log locations for JSON request bodies containing prompt/completion fields ---"
for d in /var/log/nginx /var/log/envoy /var/log/inference /var/log/vllm /opt/llm/logs; do
if [ -d "$d" ]; then
grep -rEl --include='*.log' --include='*.json' '"(prompt|messages|completion|input)"\s*:' "$d" 2>/dev/null \
| head -20 || echo " none found in $d"
fi
done
echo "--- [2] ZDR-flagged requests persisted to disk (should return NOTHING) ---"
grep -rEl '"zero_data_retention"\s*:\s*true|"zdr"\s*:\s*true|X-Zero-Data-Retention' /var/log /opt 2>/dev/null | head -20 || true
echo "--- [3] Proxy configs that log request bodies (audit these directives) ---"
grep -rEn 'log_format.*\$request_body|lua_need_request_body|capture request' /etc/nginx /etc/envoy 2>/dev/null || echo " none"
echo "--- [4] Log rotation gaps — logs older than policy retention ---"
find /var/log -name '*.log' -mtime +7 -size +10M 2>/dev/null | head -20 || true
echo "--- [5] Listening sockets not owned by expected services ---"
ss -tlnp 2>/dev/null | grep -Ev 'nginx|envoy|vllm|python|sshd|systemd' || echo " none unexpected"
echo "--- [6] Current outbound connections from serving processes (baseline for anomaly detection) ---"
ss -tnp 2>/dev/null | grep -E 'vllm|triton|python|nginx|envoy' \
| awk '{print $5, $6}' | sort | uniq -c | sort -rn | head -30 || true
echo "--- [7] Crontab/systemd timers that could indicate persistence ---"
crontab -l 2>/dev/null || true
systemctl list-timers --all 2>/dev/null | grep -vE 'systemd|logrotate|man-db|fstrim' || true
echo "=== Audit complete. Review findings above. Any hit in section [2] is a ZDR violation. ==="
} | tee "$REPORT"
Remediation
For organizations consuming AI inference (OpenRouter and similar gateways)
- Inventory your AI data flows today. Identify every application, developer workflow, and SaaS integration that sends prompts to OpenRouter or any multi-provider gateway. You cannot assess exposure to a provider breach you don't know you're routing through.
- Determine if your traffic touched the three dropped providers. OpenRouter exposes provider routing in request metadata. Query your gateway logs for the lookback window (Dragos describes months of theft) and scope which of your requests were served by the affected providers.
- Treat prompts as compromised and rotate accordingly. Any secrets, API keys, credentials, or connection strings that developers pasted into prompts routed through those providers must be rotated. This is the same guidance as any third-party breach.
- Pin providers and demand attestations. Where your gateway allows it, restrict routing to providers with published security practices, SOC 2 Type II reports, and explicit, technically enforced ZDR implementations. Ask providers the direct question: "Show me where a ZDR request body can exist on disk in your stack."
- Classify prompt content. Enforce DLP scanning on outbound LLM API traffic. Block or redact credentials, PII, and regulated data before it leaves your perimeter — your strongest control against a provider-side breach is ensuring the data never reaches the provider.
For teams operating inference infrastructure
- Enforce ZDR architecturally, not contractually. ZDR requests should be handled in-memory and never traverse a code path that writes request bodies: disable request-body logging at the proxy tier (audit nginx
log_format, Envoy access log config), disable debug/trace logging in serving frameworks, and verify analytics pipelines strip or never capture bodies. - Deploy the detections above. The defining failure in this campaign was months of dwell time with no internal detection. Egress baselining on serving hosts (script section [6]) and shell-spawn detection from serving processes are the two highest-value controls.
- Segment the data plane. Model-serving hosts should have no direct internet egress except to an allowlist of model registries and update services. Force all outbound traffic through an inspected egress proxy.
- Run the ZDR audit script across your fleet and treat any finding in section [2] as an incident, not a misconfiguration — if ZDR data is persisted, assume it is exfiltratable.
- Establish external compromise disclosure paths. All three providers learned of their breaches from outside researchers. Stand up a vulnerability disclosure program and consume third-party threat intelligence — including monitoring for your own data appearing in researcher or criminal channels.
Standing guidance
- Add AI inference providers to your third-party risk management program with the same rigor as any data processor under NIST CSF or your applicable framework.
- Include provider-side prompt theft in your incident response playbooks: define trigger conditions, rotation procedures, and customer notification obligations in advance.
- If you are subject to HIPAA or PCI-DSS, review whether prompt traffic containing regulated data transits any third-party inference provider — this disclosure is direct evidence that such providers are active intrusion targets.
Related Resources
Security Arsenal Managed SOC Services AlertMonitor Platform Book a SOC Assessment soc-mdr Intel Hub
Is your security operations ready?
Get a free SOC assessment or see how AlertMonitor cuts through alert noise with automated triage.