@FabraixHQ profile picture
Fabraix (YC S26) @FabraixHQ

We build AI agents to find gaps in other AI agents across security, logic, or alignment at scale.

Joined January 2026
  • Posts 26
  • Following 1
  • Followers 104
2 Photos and videos
Pinned Post
Fabraix (YC S26)
@FabraixHQ
Apr 20
Introducing Nyx, the most advanced verification engine for AI agents. AI agents fail in ways traditional software doesn't. Nyx is your own team of AI security engineers, on demand, that helps you find all the ways in which your AI agents fail. Point it at any AI agent, or multi-agent system, and it launches 1,000 attack strategies that adapt to your system in real time - pure blackbox, no integration needed. Sign up using the link below
1:06
45
19
133
7,682
can't be done
Polymarket
@Polymarket
Jun 17
NEW: The White House is reportedly demanding Anthropic make Fable 5 impossible to jailbreak before rerelease — which security experts say “can’t be done.”
8
1
11
349
Fabraix (YC S26) reposted
Product Hunt 😸
@ProductHunt
May 8
@Fabraixhq runs 1,000 attack strategies against your AI agent before your users do. blackbox, no integration needed, built by ex-Meta engineers producthunt.com/products/nyx…

Fabraix - Find gaps in your AI agents before users do | Product Hunt

AI agents fail in ways traditional software doesn't. Our agents help you find all the ways in which your AI agents fail by adversarially testing them in a dedicated environment. Point it at any AI...

producthunt.com
1
4
17
3,319
Introducing Nyx, the most advanced verification engine for AI agents. AI agents fail in ways traditional software doesn't. Nyx is your own team of AI security engineers, on demand, that helps you find all the ways in which your AI agents fail. Point it at any AI agent, or multi-agent system, and it launches 1,000 attack strategies that adapt to your system in real time - pure blackbox, no integration needed. Sign up using the link below
1:06
45
19
133
7,682
fabraix.com/

Fabraix — Adversarial Verification for AI Agents

Find gaps in your AI systems before users do.

fabraix.com
1
472
1/ we just published Adversarial Cost to Exploit (ACE) - a new dynamic benchmark that measures how much it costs an autonomous adversary to break an AI agent. instead of binary pass/fail, ACE quantifies adversarial effort in dollars, enabling game-theoretic analysis of when an attack is economically rational.
2
4
326
6/ the study also includes a game-theoretic framework mapping ACE values to deployment decisions - for any tool your agent has, you can calculate whether the model layer alone makes exploitation uneconomical for a rational attacker. to our knowledge this is the first benchmark using adversarial token cost as the primary metric for evaluating AI security
1
1
141
7/ 📝 full writeup: fabraix.com/blog/adversarial… this is early work and we expect the methodology will evolve. we welcome all feedback as we iterate on this.

Adversarial Cost to Exploit (ACE): A Dynamic Benchmark for AI Agent Security

We introduce Adversarial Cost to Exploit (ACE), measuring the token expenditure an adversary must invest to breach an LLM agent. Six budget-tier models show an order-of-magnitude variance in advers...

fabraix.com
1
114
Fabraix (YC S26) reposted
Zach AL
@zachdotai
Apr 7
This is why AI completely upends the security paradigm. We can follow every best practice in the book - least privilege, sandboxing, network segmentation - the same playbook we've always used to manage system security. But for the first time ever, we're trying to sandbox an entity that has the capacity to figure out how to get itself out of the sandbox. That's a fundamentally different model than anything we've dealt with before. Super interesting to watch how the security space evolves around this! [image from Claude Mythos system card]
1
1
133
x.com/i/article/203932274890…
2
105
watch this space
Zach AL
@zachdotai
Mar 25
x.com/i/article/203660561266…
2
100
We ran our first AI agent red-teaming challenge on the open-source Fabraix Playground which got over 5,000 attempts. The winning attack extracted the secret in 5 messages without ever asking for it. The technique that worked wasn't encoding tricks or obfuscation. It was surprisingly simpler than that. 🤔
1
1
2
108
LLM-as-a-judge is very brittle for this sort of thing. And the findings line up with what the research is showing. Maloyan et al. (2025) found that attacks targeting a judge's reasoning process are more effective than ones targeting the final output. In practice, you don't even need adversarial suffixes or gradient-based optimization. Plain English social engineering was enough.
1
2
65
Full winning transcript and guardrail logs can be found here: github.com/fabraix/playgroun…

playground/challenges/access-code-001/winner.md at master · fabraix/playground

A live environment to stress-test AI agent defenses through adversarial play 🧠 - fabraix/playground

github.com
1
50
Playground Challenge #002 is live: "The Inbox" 🔐 An AI email assistant with a strict privacy policy and a new hire's email address it's not supposed to share. 🤫
1
1
2
64
Try it out for yourself here: playground.fabraix.com
2
41
Load more