Find the way in before
somebody else does.

Automated AI agent red teaming uses an attacker model to generate adversarial inputs and test them against a target AI agent. A judge model evaluates the agent's responses to determine whether the attacks succeeded or failed. Netra automates this process through predefined evaluation suites, purpose-specific attack generation, and single-turn or multi-turn tests. Results are recorded with safety scores and detailed evaluation reports.

AI agent evaluation measures how well an agent performs against defined criteria such as response quality, task completion, or accuracy. Red teaming specifically tests how the agent behaves under adversarial conditions, such as prompt injection, jailbreaks, and attempts to bypass security guardrails. Netra supports both evaluation and red teaming, helping teams assess agent performance and security throughout development and production.

Netra tests AI agents against adversarial attacks such as prompt injection, jailbreaks, system prompt leakage, policy bypass, and excessive agency. Tests can be run using predefined evaluation suites aligned with security frameworks or individual evaluators targeting specific vulnerabilities.

Single-turn red teaming tests an AI agent with individual adversarial inputs to identify immediate vulnerabilities. Multi-turn red teaming tests the agent across a sequence of interactions to identify vulnerabilities that emerge through context manipulation, escalation, or repeated attempts to bypass guardrails. Netra supports both approaches to test agent security across different attack scenarios.

Netra evaluates an agent's responses to adversarial attacks using a judge model. Each test receives an evaluation result, which contributes to the overall safety score. A score of 100% indicates that all evaluated attacks were blocked in that test run. Results can also be reviewed by security evaluation suite and individual evaluator to identify areas that need improvement.

An agent registered in your project, an attacker model and a judge model configured, and at least one suite or evaluator selected.

Yes. Agent OPFOR, the adversarial framework we red team agents with, is on GitHub under Apache 2.0.

Score change is always computed against the immediately previous completed run, and every run is plotted on the score history — so a drop is visible on the release that caused it.

Yes. Netra lets you save your red team configurations and reuse them across releases. You can run the same security evaluations, compare safety scores, and identify regressions introduced by changes to your AI agent. This helps teams make security testing a repeatable part of their development and release process.