Application security company Semgrep has introduced Mandoline, a code-analysis toolkit designed specifically for AI agents, as software development teams increasingly rely on automated systems to write, review and maintain applications.
The research preview announced November 12 allows AI agents to examine how information moves through software, trace dependencies and investigate potential security weaknesses across large codebases.
Mandoline supports interfile analysis across 50 languages, including Python, Java, JavaScript, Rust, Swift, Kotlin, TypeScript and C++.
The release comes alongside Semgrep’s broader Agentic Application Security platform, which aims to integrate security checks into AI-assisted development rather than relying exclusively on traditional code reviews.
Giving AI Agents More Than Source Code
Conventional AI coding assistants can inspect files and generate security suggestions, but understanding a complex application’s behavior often requires more than reading individual code sections.
Mandoline uses static program analysis to provide structured information about call graphs, data flows, taint paths and exposed application surfaces.
For example, identifying whether an external input can reach a sensitive database operation requires tracing connections across functions, libraries and sometimes multiple files.
Instead of generating a final vulnerability verdict, Mandoline supplies code-level evidence for agents to evaluate.
The system can also communicate limitations in its analysis, including uncertainty about how particular software components interact.
Semgrep’s Drew Dennison described the technology as a collection of specialized analysis tools that agents can incorporate into their reasoning processes.
Company Reports Faster Analysis and Benchmark Results
Semgrep claims Mandoline performs relevant analyses up to 50 times faster than its existing Semgrep Pro engine.
The company also reported a 100% score on the OWASP Benchmark for Java and 98.28% for Python.
Mandoline reportedly passed more than 36,000 C++ and Java tests from the NIST Juliet suite, covering 22 vulnerability categories.
These are company-reported evaluations of specific tests, not independently verified proof that the system detects every vulnerability in production software.
The OWASP Benchmark project provides deliberately constructed vulnerable and non-vulnerable test cases for measuring detection tools’ accuracy.
Performance on such benchmarks is useful for comparing analysis capabilities but does not establish reliability across every real-world architecture or security threat.
Security Moves Into the AI Development Workflow
Mandoline also forms part of Semgrep’s new Agentic AppSec Platform, which combines Guardian, Agentic Workflows, a persistent Context Layer and integrations using the Model Context Protocol.
Guardian is intended to identify and address security problems while coding agents generate software.
According to Semgrep’s deployment findings, Guardian has reached more than 10,000 developers. The company reported that findings identified within coding-agent workflows were three times more likely to be fixed than findings raised during pull-request reviews.
It also reported a median remediation time below 30 minutes, compared with approximately one day for pull-request scans, and said 95% of Guardian scans finished within two minutes.
These performance figures are based on the vendor’s own observations.
The Emerging Competition in AI-Native Security
The development reflects a change in application security priorities.
As AI agents generate more code, developers face pressure to identify vulnerabilities without slowing down software delivery.
Semgrep’s strategy involves providing agents with deterministic analysis tools and reusable organizational context, rather than asking language models to make security judgments from source text alone.
The company reports an average 65% reduction in token costs when agents use Mandoline within its security workflows.
For engineering teams, the potential benefit is faster vulnerability investigation with clearer technical evidence.
However, automated security assessments still require testing, appropriate access controls and human oversight.
Mandoline’s significance therefore lies less in replacing security engineers than in changing the tools available to them as AI-driven software development expands.
