Skip to content
Unlisted Report logoUnlisted ReportSubscribe
AI Security

CISA Adds First AI Agent Flaw to Its Must-Patch List

LiteLLM CVE-2026-59822 lets attackers open authenticated MCP sessions with any token — the first exploited AI agent infrastructure bug in CISA's KEV.

On September 2, CISA added CVE-2026-59822, a flaw in the open-source AI proxy LiteLLM, to its Known Exploited Vulnerabilities catalog — the first confirmed, actively exploited bug in production AI agent infrastructure to make the list, Tech Insider reports.

What the flaw does

LiteLLM routes traffic between apps and large language models. The bug is in its Model Context Protocol (MCP) Streamable HTTP endpoint. An unauthenticated attacker can open a fully authenticated MCP session using any bearer token at all.

According to the official CVE record, LiteLLM's MCP Streamable HTTP endpoint "allowed an unauthenticated attacker to use a fabricated Authorization header to trigger an OAuth2 passthrough fallback path that replaced failed LiteLLM key validation with an empty UserAPIKeyAuth() object," letting requests reach MCP tooling without a valid LiteLLM key at all. The issue affects versions prior to 1.84.0 and is fixed in that release CVE.org.

It is rated high severity (CVSS 4.0 score of 8.8), with a vector indicating the flaw is remotely exploitable over the network, requires low attack complexity, and needs no privileges or user interaction CVE.org.

CISA's catalog entry was one of seven vulnerabilities added to the KEV list that day based on evidence of active exploitation, alongside flaws in products from Sangoma, SonicWall, JFrog and others CISA. Under Binding Operational Directive 26-04, federal civilian agencies are required to prioritize rapid remediation of vulnerabilities listed in the KEV catalog, particularly those that grant attackers total control of an asset after exploitation, although CISA encourages all organizations, not just federal agencies, to adopt the same risk-based patching priorities CISA.

Why it matters

MCP is the standard way AI agents connect to tools, databases and business workflows. A hijacked MCP session can let an outsider make an organization's AI agents run tools, read data or trigger actions on their behalf. As companies give agents more access, these connectors become as valuable to attackers as VPNs and email servers.

LiteLLM is maintained by BerriAI, a San Francisco startup founded in 2023 by Krrish Dholakia and Ishaan Jaffer that went through Y Combinator's Winter 2023 batch. BerriAI has raised about $1.6 million in seed funding from Y Combinator, Gravity Fund and Pioneer Fund, and reports roughly $7 million in annual recurring revenue Dealroom. The open-source LiteLLM project is widely used: it has more than 50,000 GitHub stars, supports 100-plus LLM providers through a single OpenAI-compatible interface, and is used in production by companies including NASA, Adobe, Netflix, Stripe and Nvidia, according to the company and third-party reporting Dealroom. Because it sits in front of an organization's AI infrastructure, a flaw that lets an attacker bypass authentication on its MCP endpoint effectively hands over control of whatever tools and data sources that MCP server is configured to expose — which, in production agent deployments, can include internal databases, file systems, or business APIs.

A pattern of security problems at LiteLLM

This is not LiteLLM's first security incident. According to reporting on the company's funding round, BerriAI also faced a supply-chain attack in which its PyPI package releases were compromised, and additional vulnerabilities surfaced between spring and summer 2026, which led to litigation filed against the company in a Texas federal court in April 2026 Dealroom. The repeated issues illustrate a broader challenge for fast-growing open-source infrastructure projects: as adoption scales into the hundreds of millions of daily API calls, the attack surface and the consequences of a single flaw scale with it.

Earlier AI framework warnings

It is not the first warning. In July, CISA ordered federal agencies to patch an exploited authorization flaw in Langflow, the third Langflow bug it had flagged, Anavem reported. Attackers were stealing cloud and AI service keys.

The pattern of repeated, actively exploited flaws in AI orchestration frameworks — LiteLLM, Langflow, and similar tools — reflects how quickly this category of software has been adopted in production environments, often faster than security review processes have matured around it. Since these frameworks are frequently open source and self-hosted, patching responsibility falls on the organizations running them rather than a cloud vendor applying a centralized fix, which can lead to slower real-world remediation even after a patch is available.

How the fix works

The vulnerability stemmed from an OAuth2 "passthrough fallback" code path: when LiteLLM's normal API key validation failed, instead of rejecting the request, the flawed logic substituted an empty, effectively unauthenticated user object and let the request proceed to MCP tooling anyway. BerriAI's fix in version 1.84.0 closes that fallback path, according to the GitHub security advisory and patch referenced in the official CVE record CVE.org. Organizations running LiteLLM proxy deployments with MCP routes enabled are considered at risk if they have not upgraded past that version; according to Red Hat's security advisory, products that merely use LiteLLM as a client library without running its proxy server or exposing its MCP HTTP authentication path are not affected in their default configurations Red Hat.

How MCP fits into the broader AI stack

The Model Context Protocol, originally introduced by Anthropic, has rapidly become a de facto standard for connecting AI models and agents to external tools, APIs and data sources. LiteLLM markets itself as "the AI Gateway for platform teams," offering a single key and login that spans more than 140 providers and 1,800-plus models, along with routing, spend governance, and MCP server access all behind one gateway LiteLLM. That consolidation is exactly what makes a gateway-level authentication bypass so consequential: instead of compromising one integration, an attacker who defeats the gateway's authentication can potentially reach every tool, model and MCP server an organization has connected behind it.

What teams running AI agents should do

  • Update LiteLLM to a fixed version now.
  • Keep MCP endpoints off the public internet.
  • Give each agent only the tools and data it needs.
  • Rotate model API keys and cloud credentials if you suspect exposure.

What readers should do

  • If your organization runs LiteLLM as a proxy for AI applications, confirm you are on version 1.84.0 or later, which contains the fix for CVE-2026-59822.
  • Audit whether your LiteLLM MCP Streamable HTTP endpoint is exposed to the public internet, and restrict it to internal networks or trusted clients only if so.
  • Apply the principle of least privilege to AI agents and MCP tool integrations, so that even a compromised session cannot reach more data or systems than strictly necessary.
  • Rotate API keys, model provider credentials, and any cloud service keys that may have been reachable through an exposed MCP endpoint, especially if you cannot rule out exploitation before patching.
  • Treat AI agent infrastructure — proxies, orchestration frameworks, and MCP servers — with the same security rigor as VPNs, identity providers, and email servers, since they increasingly provide similar levels of access to sensitive systems.
  • Monitor CISA's Known Exploited Vulnerabilities catalog on an ongoing basis if your organization relies on open-source AI tooling, since this is likely to be the first of several similar entries as AI agent adoption grows.
  • Even if you are not a U.S. federal agency subject to CISA's binding directive, consider adopting the same prioritization approach: patch internet-facing, actively exploited vulnerabilities first.
  • If you rely on a vendor whose product embeds LiteLLM, ask them directly whether they run the proxy server with MCP enabled, and whether they have upgraded past the vulnerable versions.

By · Published · Updated · 9 min read

Sources & Verification

Primary sources

Additional reporting

Technical sources

Read next