
AI & Agent Security
https://hmmnm.com/living-off-the-llm-attackers-weaponize-ai-infrastructure-2026/

Three AI-adjacent CVEs hit CISA’s KEV catalog in a single month: the LiteLLM proxy auth bypass (CVE-2026-42208) that exposes every conversation your org ever had with a model, a Linux kernel flaw (CVE-2026-31431) that reaches GPU training clusters, and PAN-OS exploitation (CVE-2026-0300) used as the beachhead into AI infrastructure. “Living off the LLM” is the new living-off-the-land.
In 2024, we worried about prompt injection. In 2025, we focused on AI supply chain attacks — the pattern behind the Mercor/LiteLLM breach. But 2026 has revealed something far more concerning: AI infrastructure itself has become the target. The CVE-2026-42208 vulnerability in LiteLLM — a proxy used by thousands of organizations to manage LLM API calls — demonstrated that the AI stack is now as critical as the web server stack was in 2010. For the full April picture, see our AI infrastructure under attack analysis.
This isn’t theoretical. Attackers are actively exploiting AI proxy layers, model serving infrastructure, and agentic frameworks. Here’s what every security professional needs to know about this emerging threat class.
Before diving into specific vulnerabilities, let’s map the modern AI infrastructure stack that attackers are targeting:
┌─────────────────────────────────┐
│ Application Layer │
│ (Chatbots, Agents, Copilots) │
├─────────────────────────────────┤
│ Orchestration Layer │
│ (LangChain, CrewAI, AutoGen) │
├─────────────────────────────────┤
│ Proxy/Gateway Layer │
│ (LiteLLM, OpenRouter, vLLM) │
├─────────────────────────────────┤
│ Model Serving Layer │
│ (TGI, Triton, Ollama) │
├─────────────────────────────────┤
│ Infrastructure Layer │
│ (GPU Clusters, CUDA, Drivers) │
└─────────────────────────────────┘
Each layer introduces new attack vectors that didn’t exist in traditional web infrastructure.
CVE-2026-42208 was added to CISA’s Known Exploited Vulnerabilities catalog on May 8, 2026. LiteLLM is an open-source proxy that translates requests between different LLM providers, used by organizations to:
The vulnerability allows attackers to bypass authentication and access controls, potentially gaining access to:
Think about what flows through your LLM proxy:
User Query → LiteLLM Proxy → OpenAI/Anthropic/Google
↓
Logs contain:
- Customer PII
- Internal strategy documents
- Source code snippets
- Authentication credentials
- Legal/medical/financial data
An attacker who compromises your AI proxy doesn’t just get access to your AI — they get access to everything your organization has ever asked an AI about. This is a data breach of unprecedented scale, and most organizations don’t even realize it.
Another critical vulnerability added to the KEV catalog on May 1, 2026 targets the Linux kernel. While not AI-specific, this is relevant because:
An attacker exploiting CVE-2026-31431 could:
Added to KEV on May 6, 2026, this Palo Alto Networks vulnerability reminds us that while we’re focused on AI threats, traditional network infrastructure remains under active exploitation.
The attack chain combining CVE-2026-0300 with AI-targeting techniques looks like this:
For the WordPress community, CVE-2026-41940 targeting cPanel & WHM is particularly concerning. Added to KEV on April 30, 2026 — our full breakdown is in the cPanel authentication bypass analysis — this vulnerability affects:
Given that this is a cybersecurity blog running on WordPress, here are concrete steps:
| CVE | CVSS | Target | Risk |
|---|---|---|---|
| CVE-2026-4668 | 6.5 | WordPress Amelia Plugin | SQL Injection |
| CVE-2025-71279 | 9.8 | XenForo | Passkey Authentication Bypass |
| CVE-2025-71278 | 8.8 | XenForo | OAuth2 Unauthorized Scopes |
| CVE-2026-6973 | — | EPMM | Mobile Device Management Bypass |
| CVE-2026-32202 | — | Windows | Privilege Escalation |
What these vulnerabilities collectively reveal is a fundamental shift in the threat landscape:
Traditional AppSec focuses on your code. But in the AI era, your attack surface includes:
This is the same dependency-inheritance problem we mapped in software supply chain security — one layer deeper.
When a traditional web application is compromised, the blast radius is limited to that application’s data. When an AI infrastructure component is compromised:
Your AI proxy is as sensitive as your database.
Your model weights are as valuable as your source code.
Your prompt logs are as sensitive as your email archives.
The same principles apply as in zero-trust architecture for AI systems — but stretched to cover infrastructure that answers at API speed.
Specifically for LiteLLM and similar proxies:
# Bad: Default LiteLLM configuration
litellm_settings = {
"master_key": "sk-xxx", # Single key for everything
"database_url": "postgresql://..." # Logs all conversations
}
# Good: Hardened configuration
litellm_settings = {
"master_key": None, # No master key
"auth_required": True,
"user_api_key_auth": True,
"max_budget": 100, # Limit per-user spend
"log_raw_request_response": False, # Don't log PII
"allowed_model_aliases": ["gpt-4o", "claude-3.5"], # Whitelist
}
Ask yourself:
Traditional IR plans don’t cover AI-specific scenarios. You need:
The weaponization of AI infrastructure is still in its early stages. But the addition of three AI-related CVEs to CISA’s KEV catalog in a single month signals that nation-state and criminal actors are actively targeting this space.
For cybersecurity professionals, the message is clear: your AI stack is now part of your attack surface. The tools and frameworks you adopted to accelerate development have become the very infrastructure that attackers are learning to exploit.
Stay vigilant. Patch immediately. And remember: in 2026, securing your AI infrastructure is as important as securing your web server was in 2006.
It’s the AI-era equivalent of “living off the land” attacks: instead of bringing their own malware, attackers compromise the AI infrastructure already inside your environment — LLM proxies (LiteLLM, OpenRouter), model serving stacks (TGI, Triton, Ollama), and orchestration frameworks — and weaponize its credentials, logs, and trust position.
The LiteLLM proxy authentication bypass can expose provider API keys, complete prompt/conversation history, internal documents processed through the AI pipeline, and model configurations including system prompts and fine-tuning data — effectively everything your organization has ever asked an AI about.
CVE-2026-31431 shows why the boundary is blurring: GPU clusters, model training pipelines, and Kubernetes orchestration all sit on the Linux kernel. Kernel privilege escalation on training infrastructure means access to GPU memory, model weights, and the ability to poison training runs.
Patch the three May 2026 KEV entries first (LiteLLM, kernel, PAN-OS), then harden your proxy configuration: disable the master key, require per-user auth, turn off raw request/response logging, whitelist model aliases, cap per-user budgets, and segment AI infrastructure behind zero-trust controls.
{“@context”:”https://schema.org”,”@type”:”FAQPage”,”mainEntity”:[{“@type”:”Question”,”name”:”What does Living Off the LLM mean?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”The AI-era equivalent of living-off-the-land attacks: attackers compromise the AI infrastructure already inside your environment — LLM proxies, model serving stacks, and orchestration frameworks — and weaponize its credentials, logs, and trust position.”}},{“@type”:”Question”,”name”:”What does CVE-2026-42208 expose?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”The LiteLLM proxy authentication bypass can expose provider API keys, complete prompt and conversation history, internal documents processed through the AI pipeline, and model configurations including system prompts.”}},{“@type”:”Question”,”name”:”Is a kernel CVE really an AI security issue?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”CVE-2026-31431 matters because GPU clusters, model training pipelines, and Kubernetes orchestration all sit on the Linux kernel. Kernel privilege escalation on training infrastructure means access to GPU memory, model weights, and training runs.”}},{“@type”:”Question”,”name”:”How do I start securing my AI stack today?”,”acceptedAnswer”:{“@type”:”Answer”,”text”:”Patch the three May 2026 KEV entries, then harden your proxy: disable the master key, require per-user auth, turn off raw request logging, whitelist models, cap budgets, and segment AI infrastructure behind zero-trust controls.”}}]}
Prabhu Kalyan Samal is an Application Security Consultant specializing in AI security, penetration testing, and secure architecture. He holds certifications including CompTIA SecurityX, Certified Red Team Operator, and Azure Security Engineer.
Part of our AI Agent Security: The Complete Guide (2026) series.