Endor Labs adds BYOK support for AI security scans

Endor Labs has introduced bring your own key support for AI SAST and AI Chat on its AURI platform, aimed at organisations with strict governance and data-handling requirements.

The feature lets customers route AI model requests through their own OpenAI, Azure OpenAI, Gemini or Vertex AI accounts instead of using Endor Labs' provider accounts. Where providers offer regional options, including Azure OpenAI and Vertex AI, customers can choose where requests run.

The setup is designed to address a recurring issue for companies in regulated sectors, where security reviews can slow or block tools that send source code to a vendor-selected large language model. By using an already approved provider account, organisations can keep model usage within existing procurement, compliance and audit frameworks.

The arrangement is all or nothing. Once a customer configures a key, every model request made on its behalf runs on that provider account or fails, with no fallback to Endor Labs' own credentials.

Requests still pass through Endor Labs' AI Gateway, which applies controls including encrypted key storage, provider allowlists and tenant isolation. Customers can also pass provider keys as environment variables during scans so the keys are not stored in the Endor Labs platform.

Compliance focus

The move reflects a broader push by software and security vendors to adapt generative AI tools for customers in financial services, healthcare, critical infrastructure and other tightly regulated sectors. Data residency, privacy restrictions and internal approval processes have become barriers to AI-based coding and security products that rely on third-party models.

According to Endor Labs, BYOK does not change how its AI SAST product analyses code. It still uses the same underlying models and runs in the same places, including local environments, pull requests before merge and CI/CD workflows.

The company argues that giving a general-purpose frontier model direct access to a code repository is not enough for reliable security scanning. While large language models can reason about the code they read, they often inspect only a limited portion of a codebase and can become costly when used as scanning agents across large repositories.

Endor Labs said a benchmark using the same 34 security prompts found that an agent working from a raw repository used 79.5 million tokens, while one using Endor Labs context used 6.6 million. It added that the latter completed the work 2.8 times faster.

Benchmark claims

Endor Labs also published performance figures for AI SAST against two frontier models and four traditional static application security testing tools across eight software projects written in Java, Python, JavaScript, TypeScript, C# and Go. The vendor said its security research team hand-verified the results.

According to Endor Labs, AI SAST found 192 real vulnerabilities, or 2.6 times the number found by the strongest frontier model tested and four times the high-severity findings of the strongest pattern-based SAST tool. It added that the product performed better in categories including broken access control, insecure direct object references, insecure design and business logic flaws.

Those are areas where conventional rule-based tools often struggle because they depend on how an application handles trust and authorisation rather than matching a fixed code pattern. Endor Labs said traditional scanners can flag large volumes of false positives, while frontier models can miss issues because they do not inspect enough of the codebase.

Product approach

AI SAST builds a code context graph of an application before looking for flaws, the company said. It also creates a semantic index and reads infrastructure-related files, including Dockerfiles, Kubernetes manifests and Terraform configurations, to score findings against how an application actually runs.

Detection agents then trace user-controlled data across files, functions and repositories, while triage agents classify findings and remediation agents propose code fixes. Confirmed findings include an attack vector and exploit payload as evidence, according to Endor Labs.

In one example cited by the company, its security research team used AI SAST to review open-source AI and infrastructure projects and identified a vulnerability in buffa, a Rust library created and maintained by Anthropic and scanned using Claude Code Security. Endor Labs said the issue traced untrusted input into an unbounded heap allocation and was assigned CVE-2026-55407.

BYOK support is currently available in private beta for Endor Labs customers.