Fabraix is a red-teaming service for teams shipping customer-facing AI agents, whether chat, voice, browser, or coding assistants. Its product Nyx probes a live agent from the outside without requiring access to source code, then reports the specific prompts, tool calls, and responses that produced a failure so engineers can reproduce and patch each finding. The pitch to security teams is that reading a system prompt is not enough when the underlying agent is nondeterministic and can reach untrusted files, websites, and MCP servers on its own.
Rather than cycling through a fixed list of jailbreak payloads, Nyx draws from a library of more than 10,000 attack strategies and adapts multi-turn attempts based on how the target agent defends itself. Fabraix reports a 78 percent attack success rate on the AgentHarm benchmark and says Nyx has surfaced exploitable failures in public-facing agents at dozens of Fortune 500 companies. The team has also published ACE (Adversarial Cost to Exploit), a benchmark that scores AI systems by how expensive they are for a rational attacker to break.
The company was founded in 2026 by Ahmed Aly (CEO), previously on the international payments fraud team at Monzo and the first data scientist at Sequoia-backed Two, and Ibrahim Abdu (CTO), who red-teamed AI agents inside Meta's Privacy org and built compilers for trade surveillance at TradingHub before that. Fabraix is part of Y Combinator's S26 batch, and this launch coincides with Aly publicly sharing that the company reached a $1.04M annualized run rate three months after joining YC with zero revenue.
Comments (14)
Sign in to join the discussion.
P
Priya RaghavanSep 12, 2026
Zero to seven figures in 90 days is either a masterclass in enterprise fear or a masterclass in tweet framing. Probably both.
W
wrenSep 12, 2026
What's the headcount behind this? If it's under six people I'm going to lie down for a bit.
H
Halima TraoréSep 12, 2026
Would love to chat about the customer mix. Are these mostly AI-native startups or are you seeing traditional enterprises with agent deployments already?
Ł
Łukasz K.Sep 12, 2026
The cut at 0:14 where the voiceover overlaps the terminal shot is chef's kiss. Whoever edited this earned their coffee.
M
Maya OkonkwoSep 12, 2026
How's the docs situation? Red-teaming tools live or die by whether I can wire it up in 20 minutes without pinging your Slack.
K
Kenji VasquezSep 12, 2026
Everyone's launching an AI red team this quarter. In 18 months this becomes a checkbox inside a bigger SOC platform and the standalones get squeezed.
R
Rafa M.Sep 12, 2026
Been building something adjacent for 3 weeks, about to launch in 3 more weeks, this is either very motivating or very demoralizing.
F
Farrukh IdrisovSep 12, 2026
Curious how you actually deliver findings. Is it a dashboard, a PDF, a Jira dump? Distribution of the report matters more than people think.
I
Inés BarbosaSep 12, 2026
Where do the prompts and traces from customer agents actually live? Any story on EU residency or is that a 2026 problem.
T
Tomo AlabiSep 12, 2026
Do the attack agents hit via SDK, MCP, or do I point them at an endpoint? And please tell me rate limits are configurable per target.
A
Ada PembertonSep 12, 2026
Tagline rewrite on the house: 'We break your agents on a schedule, so nobody else does it for free.' Send royalties in stickers.
D
Dima SorokinSep 12, 2026
Are you hiring offensive ML engineers? I know someone who spent two years jailbreaking frontier models for fun and is finally open to a real job.
B
Bram de VriesSep 12, 2026
Cool product. Now do continuous monitoring so I get pinged the moment someone ships a new tool call in prod.
N
Nina VolkovSep 12, 2026
The engagement on this thread is spiking way past your usual baseline. Either the ARR number did the work or someone quote-tweeted you in a big group chat.
Work With Us
Book an intro call to explore our suite of services. We might have a waitlist.
Fabraix
Our agents hack your AI, before attackers do.
Our agents hack your AI, before attackers do.
Tags
TLVC Rating
Community Rating
No ratings yet
to rate this launch.
Fabraix is a red-teaming service for teams shipping customer-facing AI agents, whether chat, voice, browser, or coding assistants. Its product Nyx probes a live agent from the outside without requiring access to source code, then reports the specific prompts, tool calls, and responses that produced a failure so engineers can reproduce and patch each finding. The pitch to security teams is that reading a system prompt is not enough when the underlying agent is nondeterministic and can reach untrusted files, websites, and MCP servers on its own. Rather than cycling through a fixed list of jailbreak payloads, Nyx draws from a library of more than 10,000 attack strategies and adapts multi-turn attempts based on how the target agent defends itself. Fabraix reports a 78 percent attack success rate on the AgentHarm benchmark and says Nyx has surfaced exploitable failures in public-facing agents at dozens of Fortune 500 companies. The team has also published ACE (Adversarial Cost to Exploit), a benchmark that scores AI systems by how expensive they are for a rational attacker to break. The company was founded in 2026 by Ahmed Aly (CEO), previously on the international payments fraud team at Monzo and the first data scientist at Sequoia-backed Two, and Ibrahim Abdu (CTO), who red-teamed AI agents inside Meta's Privacy org and built compilers for trade surveillance at TradingHub before that. Fabraix is part of Y Combinator's S26 batch, and this launch coincides with Aly publicly sharing that the company reached a $1.04M annualized run rate three months after joining YC with zero revenue.
Comments (14)
Zero to seven figures in 90 days is either a masterclass in enterprise fear or a masterclass in tweet framing. Probably both.
What's the headcount behind this? If it's under six people I'm going to lie down for a bit.
Would love to chat about the customer mix. Are these mostly AI-native startups or are you seeing traditional enterprises with agent deployments already?
The cut at 0:14 where the voiceover overlaps the terminal shot is chef's kiss. Whoever edited this earned their coffee.
How's the docs situation? Red-teaming tools live or die by whether I can wire it up in 20 minutes without pinging your Slack.
Everyone's launching an AI red team this quarter. In 18 months this becomes a checkbox inside a bigger SOC platform and the standalones get squeezed.
Been building something adjacent for 3 weeks, about to launch in 3 more weeks, this is either very motivating or very demoralizing.
Curious how you actually deliver findings. Is it a dashboard, a PDF, a Jira dump? Distribution of the report matters more than people think.
Where do the prompts and traces from customer agents actually live? Any story on EU residency or is that a 2026 problem.
Do the attack agents hit via SDK, MCP, or do I point them at an endpoint? And please tell me rate limits are configurable per target.
Tagline rewrite on the house: 'We break your agents on a schedule, so nobody else does it for free.' Send royalties in stickers.
Are you hiring offensive ML engineers? I know someone who spent two years jailbreaking frontier models for fun and is finally open to a real job.
Cool product. Now do continuous monitoring so I get pinged the moment someone ships a new tool call in prod.
The engagement on this thread is spiking way past your usual baseline. Either the ARR number did the work or someone quote-tweeted you in a big group chat.