Two critical flaws in one of the most widely-used LLM gateways, LiteLLM, could let low-privilege users execute code, hijack admin accounts, and steal AI provider API keys.

With 97 million downloads per month and more than 59,000 GitHub stars, LiteLLM runs inside large enterprises, startups and many of the platforms used to build AI applications.

It functions as an LLM gateway, used for storing API keys for every AI provider the organization uses, managing user access through roles and permissions, routing prompts and responses between applications and providers and tracking usage and costs across models.

However, according to Checkpoint researchers, CVE-2026-35029 allowed any user – including those with read only access – to reach administrative functions and take over the gateway. CVE-2026-59823, meanwhile, allowed attackers to redirect gateway requests to their own server and capture AI provider API keys.

“In practical terms, anyone holding an account on a vulnerable LiteLLM deployment could take over the gateway and use it for malicious purposes,” said the firm.

LiteLLM uses role-based access control, where administrators manage the system and other users receive limited or read-only access. However, the CVE-2026-35029 access control vulnerability means that users with minimal privileges could reach administrative functions controlling the gateway’s core configuration.

From there, an attacker could read files stored on the server, including configuration files and system files containing sensitive information. They could also overwrite administrator credentials with their own, locking legitimate administrators out of the system, and instruct the gateway to download and execute code from an external source, enabling malware installation, data exfiltration or movement to other systems on the network.

CVE-2026-59823 is a server side request forgery vulnerability that allowed any gateway user to direct LiteLLM to send API requests to a server under the attacker’s control without validation.

Attackers could steal AI provider API keys without any compromise of OpenAI, Anthropic or other providers, use the organization’s AI provider accounts without authorization and cause unexpected costs to be billed to the organization.

LiteLLM has now resolved CVE-2026-35029 and CVE-2026-59823 in version 1.83.9, with all later versions including the fixes.

It’s notable that the two vulnerabilities are nothing new: they include missing authorization checks, path traversal, RCE via dynamic module loading, and insecure password storage, and all have established mitigation patterns.

“Both vulnerabilities belong to well understood classes of security error, which shows that AI infrastructure is now exposed to the same risks that enterprise software has faced for years,” warned Checkpoint.

“As tools like LiteLLM gain adoption at speed, attackers are paying close attention to them, and organizations need to secure their AI gateways with the same rigor they apply to any other business critical system.”

AI gateways are starting to emerge as high-value targets, holding credentials, managing trust relationships, and controlling access to multiple AI services, said Li Zhao, principal strategic services consultant at Black Duck.

“Because many organisations consume these technologies indirectly through platforms, frameworks, and internal tooling, the risk extends beyond a single product and becomes a broader ecosystem concern,” she said.

“Organizations should treat AI gateways with the same level of scrutiny they apply to identity platforms, API gateways, secrets-management systems, and other high-trust enterprise infrastructure.”