Back to Articles

EASM

Your AI Stack Is an Attack Surface: Langflow, LiteLLM, MLflow and the AI Act from August 2026

12 min read

In 2026 the CISA KEV catalog filled up with AI tooling: Langflow, LiteLLM, MLflow, Ray, n8n, Marimo and GPUStack. What was exploited, why these services hold your keys, how to harden them, and what the AI Act asks of you after the Omnibus.

ai-security
attack-surface

By November 2026 the pattern was hard to miss. In March, CISA added n8n and Langflow to its Known Exploited Vulnerabilities (KEV) catalog; by early November the list of AI and agent tooling under active exploitation also included Marimo, LiteLLM, Ray, MLflow and GPUStack. The workflow builders, LLM gateways and machine learning platforms that companies stood up during their AI pilots have become part of the external attack surface, and attackers treat them exactly that way. This article reconstructs what was exploited during 2026, explains why these services are such valuable targets, and sets out the hardening that closes most of the gap. It is a companion to our guide to external attack surface management, because the first problem is almost always visibility: you cannot protect an AI service that nobody told you was online.

Key Takeaways
  • Between March and November 2026, CISA added 13 vulnerabilities in AI and agent tooling to the KEV catalog: five in Langflow, three in LiteLLM, and one each in n8n, Marimo, Ray, MLflow and GPUStack.
  • These services are high-value targets because they hold provider API keys and cloud credentials, and because many of them execute code by design.
  • According to the Cloud Security Alliance, the Langflow flaw CVE-2026-33017 was exploited about 20 hours after disclosure.
  • IBM's 2026 breach study found that more than 20% of organisations reported a breach targeting AI models or applications, most often through compromised APIs, apps or plug-ins and cloud misconfigurations.
  • After the Digital Omnibus on AI, the AI Act's transparency duties still apply from 2 August 2026, but it does not tell you to patch Langflow: securing the infrastructure falls under NIS2 and the GDPR.

AI tooling in the KEV catalog: the 2026 list

The KEV catalog only lists vulnerabilities for which CISA has evidence of exploitation in the wild. That makes it a useful, conservative signal: an entry is not a theoretical risk but something attackers already used. The table below collects every entry for AI, LLM and agent tooling added between March and early November 2026, with the description CISA published.

Added to KEVProductCVEWhat it allows (per CISA)
11 Marchn8nCVE-2025-68613Remote code execution through the workflow expression evaluation system
25 MarchLangflowCVE-2026-33017Code injection that allows building public flows without authentication
23 AprilMarimoCVE-2026-39987Pre-authentication remote code execution with shell access
8 MayLiteLLMCVE-2026-42208SQL injection that reads, and can modify, the proxy database and the credentials it manages
21 MayLangflowCVE-2025-34291Permissive CORS plus a SameSite=None refresh cookie: a malicious web page obtains tokens and can reach code execution
8 JuneLiteLLMCVE-2026-42271Command injection on the host by any authenticated user, including low-privilege internal keys
7 JulyLangflowCVE-2026-55255An authenticated user can run any other user's flow by supplying its ID
21 JulyLangflowCVE-2026-0770Remote code execution through functionality included from an untrusted source
4 AugustLangflow (listed under IBM)CVE-2026-9198Unauthenticated remote code execution on default deployments
17 AugustRayCVE-2025-62593Code injection leading to remote code execution, reachable through Firefox and Safari when Ray is used as a development tool
19 AugustMLflowCVE-2026-64849Server-side request forgery towards internal services and cloud metadata, with the response returned
2 SeptemberLiteLLMCVE-2026-59822Authentication bypass on the MCP Streamable HTTP endpoint using an arbitrary Bearer token
9 NovemberGPUStackCVE-2026-102895Unauthenticated administrative access, exposure of worker and cloud-provider secrets, code execution

Two things stand out. First, the speed. According to a research note from the Cloud Security Alliance, CVE-2026-33017 in Langflow was exploited roughly 20 hours after it was disclosed. A monthly patch cycle does not survive that timeline. Second, the repetition: Langflow appears five times in eight months, LiteLLM three times. These are young, fast-moving open-source projects, and the same product keeps coming back. If you run one of them, assume you will need to update it again soon.

If you want to see where these entries sit within the whole year of exploited vulnerabilities, our analysis of the 2026 KEV entries puts them in context alongside firewalls, VPNs and collaboration servers.

Why these services hold the keys

An exposed web server with a known flaw is a problem. An exposed AI service with a known flaw is often a bigger one, because of what sits inside it.

An LLM gateway is a vault of API keys

Tools such as LiteLLM act as a proxy between your applications and model providers. To do that, they store the provider API keys, issue their own virtual keys to internal teams, and keep usage records in a database. That is why CISA's description of CVE-2026-42208 matters: an SQL injection that reads the proxy database gives access to "the credentials it manages". A compromised gateway can mean stolen provider keys, unexpected bills, access to every prompt and response routed through it, and a foothold in the network where it runs.

Orchestration tools run code by design

Langflow, n8n and notebook environments such as Marimo exist to let users build flows that call APIs, run scripts and connect to data. Code execution is a feature, protected by authentication. When the authentication fails or can be bypassed, as in several of the entries above, the attacker inherits a ready-made remote execution engine, usually with network access to internal systems and stored connection secrets.

ML platforms and GPU clusters lead into the cloud account

MLflow, Ray and GPUStack typically run in cloud environments next to training data, model registries and object storage. The MLflow SSRF in the KEV catalog reaches cloud metadata services, the classic route to temporary cloud credentials. The GPUStack entry explicitly mentions the disclosure of worker and cloud-provider secrets. In both cases the vulnerable service is a door into the wider cloud account.

What attackers do: the Wiz honeypots

The most concrete view of attacker behaviour in 2026 came from Wiz Threat Research, which ran honeypots imitating AI services for 90 days. The Cloud Security Alliance summarised the study on 1 September 2026. The decoys mimicked LiteLLM, Flowise, LangChain, Langflow, ChromaDB, Ollama, OpenWebUI and Node-RED. Among the activity reported:

  • Malicious MCP configurations. Attackers submitted MCP stdio server configurations that launched cryptominers, disguised as ordinary processes, while still answering the protocol handshake correctly.
  • Bypassing LiteLLM authentication. Fabricated Bearer tokens, in some cases just the letter x, were used against the flaw later tracked as CVE-2026-59822 to reach MCP tooling without a valid key.
  • Blind prompt injection. Payloads sent to agent frameworks were confirmed through out-of-band DNS callbacks, a sign that attackers test whether an agent will act on injected instructions even when they cannot see its output.
  • Credential theft from memory. Attackers read the LiteLLM master key directly from the running process.

The note also records what was not seen: no attacks specific to Ollama or ChromaDB were observed in the period. Its recommendations are practical: update LiteLLM to 1.83.7 or later, make sure malformed or single-character Bearer tokens are rejected, keep MCP stdio command endpoints away from public reach and limited to administrators, rotate the provider keys issued through the proxy if exposure cannot be ruled out rather than relying on log review alone, require authentication by default, restrict outbound traffic and IAM permissions, and watch for unusual child processes under AI framework directories.

Shadow AI: the pilot on a public IP

Most of the exposed AI services we would worry about were not deployed by the security team. They were proofs of concept: a data scientist launches Langflow on a cloud VM to show a demo, a developer puts a LiteLLM proxy in front of three providers to compare costs, an innovation team spins up a GPU node for a hackathon. The service gets a public IP because that is the quickest way to share it, it runs with default settings, and nobody adds it to the asset inventory. It is the same mechanism we described in our article on shadow IT, with one difference: these pilots usually hold real API keys and connect to real data.

IBM's Cost of a Data Breach Report 2026 gives a sense of scale. According to IBM's announcement, more than 20% of organisations reported a breach that targeted AI models or applications. The most common causes were not exotic model attacks but ordinary weaknesses in the systems around them: compromised APIs, applications or plug-ins (27%) and cloud misconfigurations affecting AI workloads (27%). We cover the rest of the report in our article on the cost of a data breach in 2026.

The practical consequence: before you can harden your AI stack, you need an inventory of it, and the inventory has to be built from the outside as well as from the inside. Ask your teams what they deployed, and also check what actually answers on your IP ranges and cloud accounts.

Hardening: five controls that close most of the gap

1. Authentication on everything, starting with defaults

Several of these tools have shipped, at some point, with authentication optional or disabled in the default configuration, or with public endpoints that bypass it. Check every instance: login enforced, no anonymous or public flows unless you deliberately want them, default and demo accounts removed, administrative APIs separated from user APIs. For gateways, reject malformed tokens and give each team its own scoped key rather than sharing the master key.

2. A private network, not a public IP

An orchestration tool or an LLM gateway rarely needs to be reachable from the whole internet. Put it behind a VPN, a zero-trust access proxy or at least an IP allowlist. If an application must call it from outside, expose only that application's API through a reverse proxy with authentication, never the admin interface. On cloud platforms, check security groups and firewall rules for ports opened to 0.0.0.0/0 during the pilot and never closed.

3. Secrets out of the application

Provider API keys, database passwords and cloud credentials should live in a secrets manager, not in environment files on the VM or in the tool's own database. Give each key the narrowest scope and the lowest spending limit that works, and plan for rotation: if a vulnerable version was exposed, rotate the keys it held, even if the logs look clean. On cloud hosts, require IMDSv2 or the equivalent protection for the metadata service, which blunts SSRF flaws like the MLflow one.

4. Egress control and process monitoring

Many of the attacks above end with the server downloading a cryptominer or a second-stage payload, or calling back to an attacker's domain. Restricting outbound traffic to the model providers and services the tool actually needs breaks that chain. Monitoring which processes the AI service spawns, and alerting on shells or unexpected binaries under its directories, catches what gets through.

5. Versions you track against KEV

These projects release often and are exploited quickly. Keep an inventory of which version of each tool runs where, subscribe to the GitHub security advisories of the projects you use, and treat a KEV entry for your stack as a same-day priority. Our article on prioritising vulnerabilities with KEV, EPSS and SSVC explains how to fit this into a wider process without drowning in alerts.

What the AI Act does and does not require from 2 August 2026

Many companies expected 2 August 2026 to be the big AI Act deadline. The Digital Omnibus on AI changed that. According to the summaries published by law firms such as Lewis Silkin and Cuatrecasas, the amending regulation, Regulation (EU) 2026/1744, was published in the Official Journal on 24 July 2026 and entered into force on 27 July 2026. The main dates, as those summaries describe them:

ObligationApplies from
Transparency duties under Article 50, such as informing people that they are interacting with an AI system2 August 2026 (unchanged)
Machine-readable marking of generative AI output, for systems placed on the market before 2 August 20262 December 2026
Stand-alone high-risk AI systems listed in Annex III2 December 2027 (deferred from 2 August 2026)
High-risk AI embedded in products regulated under Annex I2 August 2028
AI literacy (Article 4)Softened: providers and deployers must take measures to support it, without a mandated "sufficient level" for each person

Check the final text in the Official Journal before you rely on any single date: these are summaries, and your obligations depend on your role as provider or deployer and on the systems you use. Our AI Act guide (PDF) covers roles and risk classes in more depth.

The important point for IT and security teams is what the AI Act does not do. It does not tell you to patch Langflow, close a LiteLLM port or rotate an API key. The security of the infrastructure that runs your AI tools is governed by the rules you already know: if you are in scope of NIS2, the risk-management measures of Article 21 cover vulnerability handling, access control and supply chain security for these systems like any other; and if the tools process personal data, Article 32 of the GDPR requires security appropriate to the risk. A breach of an exposed AI gateway is reported and judged under those regimes, not under the AI Act. Delaying the high-risk obligations to 2027 does not delay anything on this front.

Checklist: securing AI infrastructure

  1. Build the inventory. List every AI tool in use (workflow builders, LLM gateways, notebooks, model registries, vector stores, GPU clusters, MCP servers), who owns it, where it runs and which version it is.
  2. Check external exposure. Scan your domains, IP ranges and cloud accounts for AI services answering from the internet, including those nobody declared.
  3. Close the KEV list. Confirm that no instance of n8n, Langflow, Marimo, LiteLLM, Ray, MLflow or GPUStack runs a version affected by the entries in the table above.
  4. Enforce authentication. No anonymous access, no public flows by default, separate admin interfaces, malformed tokens rejected.
  5. Move admin interfaces off the internet. VPN, access proxy or allowlist; no 0.0.0.0/0 rules left from pilots.
  6. Move secrets to a vault and rotate exposed ones. Provider keys with narrow scopes and spending limits; rotation after any exposure of a vulnerable version.
  7. Restrict egress and protect metadata. Outbound traffic only to required services; IMDSv2 or equivalent on cloud hosts.
  8. Lock down MCP. Stdio command endpoints only for administrators and never public; review which tools each agent can call.
  9. Monitor processes and logs. Alert on shells and unexpected binaries spawned by AI services; keep prompt and access logs where they are needed for investigation.
  10. Map the obligations. Record which AI Act duties apply to you from 2 August 2026 and handle the infrastructure risk inside your NIS2 and GDPR security measures.

How Orizon can help

Most exposed AI services are found the same way an attacker finds them: from the outside. RECON External starts from your domain name and finds every subdomain, port and service you expose to the internet, including the ones you did not know about. It fingerprints the servers, frameworks and products it finds, recognises login and admin panels by their favicon even on unusual ports, matches products to known CVEs weighted by CISA KEV and EPSS, and then tests each finding from an isolated server, marking it Proven, To verify or Closed with the evidence. For AI workloads in cloud accounts you connect through read-only connectors, RECON Cloud adds AI Security Posture Management: inference endpoints, vector stores, model registries and MCP servers open without authentication, untrusted or pickle-format models, over-privileged agent tools, sensitive training data and shadow AI, mapped to the OWASP Top 10 for LLM Applications and the NIST AI RMF. For the broader method behind continuous discovery, return to our EASM guide.

ai-security
attack-surface
kev
langflow
litellm
mlflow
mcp
shadow-ai
ai-act