ZeroLeaks

Overview

ZeroLeaks provides security testing for AI agents and LLM system prompts. Its hosted product continuously red-teams agents through tools, memory, credentials, MCP servers, permissions, and connected agents, while the source-available CLI and TypeScript library test standalone system prompts for extraction and injection vulnerabilities.

Key Features

• Hosted continuous red teaming tests agents on every change, including tools, memory, MCP servers, permissions, and connected agents.
• Automated CI/CD scans can start from a pipeline through the REST API or SDK and gate releases on the returned score.
• Each finding is ranked by severity and includes attack evidence, affected behavior, and remediation; the next scan reruns the attack to confirm the fix.
• Shield flags prompt injection in user messages, retrieved documents, and tool results with one API call.
• The open-source CLI and TypeScript library run extraction and injection scans on a system prompt using selected model providers.
• The probe library draws on published attacks and sources, with an injection corpus split across extraction, tool hijacking, indirect injection, authority abuse, multi-turn grooming, and protocol exploits.

Use Cases

• Security teams red-teaming deployed AI agents before launch and on every change.
• Developers testing standalone system prompts for prompt extraction and injection vulnerabilities.
• Teams adding automated security gates to CI/CD pipelines for agent releases.
• Applications checking user messages, retrieved documents, and tool results for prompt injection with Shield.

Getting Started

• For hosted use, start a 14-day trial or link an agent endpoint and run a scan from the dashboard.
• For the open-source scanner, install with npm install zeroleaks or bun add zeroleaks.
• Set OPENROUTER_API_KEY or another supported provider configuration, then run zeroleaks scan --prompt followed by the system prompt.
• Programmatic use can call runSecurityScan with a system prompt and model options.

Deployment & Requirements

• The open-source scanner requires at least one model-provider key or an OpenAI-compatible base URL for a local server.
• Hosted agent scans target a configured HTTP endpoint or run through your own agent code with the SDK.
• Enterprise plans offer a dedicated private-cloud deployment; other plans are hosted.

Before You Adopt

• License: Other. Review its terms before using, modifying, or distributing the project.
• Running the open-source scanner requires configuring model-provider credentials or an OpenAI-compatible local endpoint.
• The hosted API no longer runs standalone prompt scans; that capability is handled by the CLI package.
• The hosted service retains certain conversation logs, tool logs, skill archives, and runtime-scan instructions for up to 365 days, or less on higher plans.
• Shield request logs normally store metadata rather than request text, with a redacted excerpt only when the free Shield model flags a request on a free plan.