# AgenticBench

> Independent security leaderboard for AI coding agents and harnesses

- Website: https://agenticbench.org
- Category: Monitoring & observability
- Launched on RankRoot: 2026-10-27
- Visitors sent by RankRoot: 0
- Show HN discussion: https://news.ycombinator.com/item?id=50262290

## About

AgenticBench is an independent benchmark that tests AI coding agents and harnesses for machine and data security. It evaluates how agents handle egress, telemetry, workspace boundaries, and secrets in a fresh container. Run by Agentic Thinking, the project takes no vendor money and provides rolling test results.

## Who it's for

Developers and companies evaluating the security of AI coding agents and harnesses.

## Use cases

- Checking security test results for specific AI coding agent versions
- Comparing data egress and workspace boundaries across different coding harnesses
- Reviewing vendor disclosures and security test failures for AI coding tools

## Features

- Rolling leaderboard of AI coding agent security tests
- Evaluates system-call traces, model requests, and network traffic
- Tests 9 security parameters including secrets handling and restricted modes
- Independent research lab testing with no vendor financing

## FAQ

### What does AgenticBench do?

AgenticBench is an independent benchmark that tests AI coding agents and harnesses to measure how they handle machine security, data egress, telemetry, workspace boundaries, and secrets.

### Who is AgenticBench for?

It is for developers and companies who want to verify whether an AI coding agent respects controls, privacy, and security boundaries before allowing it near source code and secrets.

### How many security tests does AgenticBench run?

AgenticBench runs 9 security tests on each tested AI coding agent version in a fresh container.

### Who runs AgenticBench?

AgenticBench is run by Agentic Thinking, an independent AI agent research lab in the UK that takes no vendor money and never scores its own software.


Source: https://getrankroot.com/launches/agenticbench. Product facts come from the maker or the product's own website.