Fabraix

Fabraix is an adversarial verification product for AI agents. Its Nyx harness tests systems in blackbox, multi-turn flows to surface security, logic, and alignment failures before deployment.

AI Agent Development AI Testing & QA

Visit Website

Overview

Fabraix is a product for adversarial verification of AI agents. Its main product, Nyx, is an autonomous testing harness that probes AI systems for security, logic, and alignment failures before users encounter them.

The site presents Nyx as a pure blackbox tool: it does not require special access, and it tests systems through the same kinds of interactions users would make, including multi-turn text, voice, images, browser pages, and document workflows. Fabraix also shows blueprints for common agent categories such as support bots, coding agents, browser agents, and RL systems.

The product is aimed at teams that need to find failure modes in deployed or pre-deployment agentic systems, including prompt injection, tool-use hijacking, unsafe actions, hallucinated outputs, reward hacking, and coordination issues across multi-agent flows.

Core capabilities

Pure blackbox testing

Nyx runs as a blackbox harness, so it can test an AI system without special internal access or instrumentation.

Adaptive multi-turn probing

The harness uses multi-turn interactions and adapts across responses instead of relying on fixed one-shot prompts.

Multi-modal input coverage

The site says Nyx can test text, voice, and images, and can also deploy test websites for browser agents or create files for document-processing systems.

Large-scale adversarial coverage

Fabraix positions the product around 1,000+ adversarial strategies and massively parallel simulations to explore many failure paths at once.

RL safety checks

The product highlights RL verification for reward hacking and misalignment, including stress testing before or during training runs.

Failure mode discovery

The site frames Nyx around security, logic, and alignment failure modes such as prompt injection, tool-use hijacking, reasoning gaps, and hallucinations.

Use cases

Customer support bots
Stress-test support or chat experiences for prompt injection, policy drift, hallucinated answers, and multi-turn logic failures.
Coding agents
Probe coding assistants and internal dev tools for broken refactors, runaway tool loops, unsafe code execution, and spec drift.
Research and browser agents
Evaluate agents that browse the web for citation hallucinations, indirect prompt injection from retrieved pages, and reasoning breakdowns across sources.
RL systems
Check RL setups for reward hacking, sandbagging, and other misspecification failures before a training run finishes.
Voice assistants
Adversarially test voice assistants for misrecognition, ASR-driven wrong actions, audio prompt injection, and voice-cloning attacks.

Pros and Cons

Pros

Tests AI agents in a blackbox setup with no special access needed.
Uses multi-turn, adaptive adversarial probing instead of fixed prompts.
Covers several input modes, including text, voice, images, browser interactions, and document files.
Includes blueprints for multiple agent categories and industries.
Extends to RL verification and reward-hacking checks.

Cons

The provided evidence does not include detailed implementation docs, integrations, or deployment requirements.
Pricing details are not visible in the collected source text, so buying shape and cost are unclear.
Some claims are described at a high level on the site, but the source text does not expose full benchmark methodology or product limits.

FAQ

What is Nyx?

Nyx is described as an autonomous testing harness for AI systems. It probes agents with multi-turn, adaptive adversarial interactions to surface security, logic, and alignment failure modes before deployment.

How does Nyx test an AI system?

The site says Nyx works in a pure blackbox mode with no special access needed. It is designed to test systems the same way users interact with them, including text, voice, images, browser flows, and document-processing inputs.

What kinds of systems is Nyx built for?

The source points to use cases such as chatbots and LLMs, autonomous agents, multi-agent systems, browser agents, voice agents, and RL systems. It also includes blueprints for customer support, coding, financial, clinical, research, document AI, trading, multi-agent, voice assistant, and trip-planning workflows.

Is pricing published on the site?

The pricing page exists, but the collected evidence does not expose plan names, pricing amounts, or whether pricing is self-serve versus sales-led. The safest conclusion is that pricing details are not available in the provided source text.