About
Fabraix builds adversarial red-teaming AI agents, including its autonomous Nyx agent, for companies deploying customer-facing chat, voice, browser, and coding agents. Its black-box, massively parallel testing surfaces security, logic, and alignment exploit paths that static benchmarks and one-off audits can miss, while its ACE benchmark adds a game-theoretic measure of attacker cost.
Market
Fabraix competes in AI security testing and adversarial verification for customer-facing AI agents, a segment spanning LLM evaluation, AI red teaming, application security, and agent reliability. Its positioning is an autonomous, managed offensive-security service: unlike checklist-oriented tools, Nyx operates in pure black box, adapts across hundreds of turns, chains exploits, tests multiple interaction surfaces, and returns prioritized findings with reproduction and remediation context.
Fabraix primarily targets enterprise and growth-stage companies that ship customer-facing chat, voice, browser, RAG, email, or coding agents—especially organizations with high-value or regulated workloads. The likely buyers are AI platform, application security, product security, and engineering leaders responsible for validating agent behavior and preventing exploitable failures before deployment or at scale.
At a Glance
Problem
AI agents break in ways traditional software does not: changing a model, prompt, tool, permission, or data source can alter behavior even when application code remains unchanged. For companies deploying customer-facing chat, voice, browser, or coding agents, manual red-teaming is slow, costly, incomplete, and quickly becomes stale after each release. Fabraix frames the economic pain as a recurring security-testing problem: a manual engagement can take weeks and cost six figures, while security teams lack the bandwidth to retest agents at the pace of deployment.
The killer use case is continuously stress-testing a production or pre-production customer-facing agent before changes ship, finding jailbreaks, prompt injections, data-exfiltration paths, logic failures, and alignment failures that ordinary benchmarks or one-off audits miss. The value is particularly high for organizations whose agents can access tools, websites, files, messages, or sensitive business workflows.
Product / Service
Fabraix’s product is Nyx, an autonomous, black-box red-teaming and adversarial-verification harness. It interacts with an agent directly and indirectly through its environment, supports text, voice, images, browser interactions, and document inputs, and does not require source code, model weights, network access, or a special integration. Nyx starts from a library of more than 10,000 attack strategies and jailbreaks, then adapts over multi-turn campaigns and uses adversarial self-play so successful attacks generate stronger variants.
The delivery model spans free access for approved non-commercial researchers, one-off scans priced per target, continuous testing priced monthly by usage, and enterprise deployments with dedicated infrastructure, SSO, audit logging, custom SLAs, policy coverage, and remediation support. Findings include reproduction steps and remediation guidance, while CI/CD and scheduled scans can retest every prompt, model, tool, or agent change. Fabraix says Nyx typically finds first exploits in under an hour, covers more ground than a manual audit at a fraction of the cost, and keeps testing current as the agent changes.
Market
Fabraix operates in AI security testing, AI red teaming, and offensive verification for LLM applications and autonomous agents. Its competitive set includes open-source or developer-oriented red-team tooling such as Promptfoo, Giskard’s continuous red teaming, and Mindgard’s automated AI-lifecycle testing, alongside broader AI-security platforms such as Lakera. Nyx differentiates around autonomous adaptive attacks, black-box multimodal and indirect environmental testing, self-improving attack generation, and continuous CI/CD coverage rather than static test suites or point-in-time assessments.
The company is an early-stage YC Summer 2026 startup founded in 2026 with a two-person team, so the available evidence does not establish recurring revenue or a paid-customer count and does not justify calling it definitively pre-revenue. It does, however, report meaningful early traction: Nyx has found vulnerabilities in agents operated by dozens of Fortune 500 companies, and it achieved a 78% attack-success rate on the AgentHarm benchmark versus 67% for GPT-5.6 Sol. Fabraix publicly launched the product in July 2026 and is soliciting demos and scans, indicating an early commercial rollout rather than a mature scaled business.
Founders & Leadership
Funding History
BYLD (Dubai), Y Combinator
Recent News
Y Combinator’s security-startup directory listed Fabraix as a 2026-funded company, describing its AI red-teaming agents for continuously detecting vulnerabilities in customer-facing AI.
A roundup identified Fabraix as a company founded by Ahmed Aly and Ibrahim Abdu that develops AI red-teaming agents to continuously detect security vulnerabilities in customer-facing AI.
Fabraix’s Y Combinator launch announcement introduced its AI red-teaming agents and said the product had already found vulnerabilities in agents at dozens of Fortune 500 companies.
Fabraix’s case-studies page highlighted production-agent vulnerabilities—including sandbox escapes, secret exfiltration, and discriminatory steering—identified by its Nyx red-teaming product.
Y Combinator’s company profile described Fabraix as a Summer 2026 startup building AI red-teaming agents. Its product, Nyx, continuously tests customer-facing AI across chat, voice, browser, and coding environments.
Active Roles
0No active roles right now.
Get notified when they postBusiness Model
Fabraix charges customers per adversarial scan or for continuous testing through CI. It also offers enterprise-oriented capabilities such as dedicated infrastructure, SSO, audit logging, custom SLAs, and remediation support, while providing free access for non-commercial researchers.