Ammar (@aamarsbarr) header
Ammar (@aamarsbarr) profile picture

building robots for the future | ML/AI @Autodesk | Private Pilot | Tron Eng

Toronto 🇨🇦
Joined October 2025
Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)Photo by Ammar (@aamarsbarr)
Poseidon just beat GPT-6 Astra by 17.6 points and Claude Fable 5.2 by 2.9 on CWE-Bench, a benchmark built for one job: finding, proving, and fixing vulnerabilities in real software. A specialized model beating two frontier generalists at their own sub-task used to be rare. In security right now it's the whole business model. Horizon3.ai tripled its valuation to $2B in 14 months on a $250M Series E, running what it calls AI Hackers across 310,000 production tests with zero outages and close to $100M in ARR. XBOW raised $237M, passed a $1B valuation, and its autonomous agent sits at #1 on HackerOne's leaderboard with over 1,060 submissions. RunSybil picked up a $40M seed from ex-OpenAI and ex-Meta red teamers. Pentera is past $100M ARR with 1,200 customers. Terra Security and Hadrian are chasing the same Fortune 500 budget line. The benchmarks tell a messier story than the funding rounds do. The best agent on CVE-Bench solves 13 percent of real CVEs. AutoPenBench drops from 64 percent with human hints to 21 percent without them. GPT-4 exploits 87 percent of one-day CVEs when handed the advisory text, and 7 percent when it has to find the bug itself. HackerOne still paid out $81M to human hackers last year, up 13 percent, even as valid AI-sourced reports rose 210 percent. The money is chasing speed to discovery, and those benchmark gaps say model quality isn't the thing holding it back. The same week Poseidon posts a benchmark win, Google is separately describing an unattributed framework running this exact skill on the offense side, for free, off a staging server, with no term sheet and no HackerOne leaderboard. Both sides are automating the identical capability. Only one side needed to raise money to do it. teapawt.net/thejunaidm/status/2122…
Introducing Poseidon: a state of the art cybersecurity specific model, designed for vulnerability research, finding, prove, and fix vulnerabilities in real software. Poseidon preforms: - 17.6% higher than GPT-6 Astra - 2.9% higher than Claude Fable 5.2 (CWE-Bench)
1
1
5
273
Ammar retweeted
6 months ago Trident started post-training Poseidon because nothing else worked. frontier models were completely gated on offensive security and none of the cyber programs existed yet. It was difficult and at some point our team thought we were wasting our time in the long run. This is because we assumed once major labs rolled out programs like Daybreak and CVP, a custom cyber-model wouldn't make sense anymore. Now that those programs are out, it's pretty clear we had it backwards; general-purpose frontier models still carry bloated inference costs, overhead and aren't accessible to everyone which makes continuous security analysis at scale impractical Our team post-trained Poseidon on hundreds of thousands of CVEs and vulnerability samples alongside non-public artifacts, pre- and post-patch diffs, stripped binaries, raw crash dumps, and execution traces. We removed non-domain weights, compressed and quantized it for faster inference, and stripped refusal guardrails at the weight level Because it's built strictly for security, it wins on inference cost and speed while performing better than frontier models at auditing large codebases, reversing stripped binaries, triaging fuzzer crashes, validating exploitability etc.. Poseidon is now the core model powering Trident’s offensive capabilities and reasoning.
1
1
35
3,940
Ammar retweeted
Introducing Poseidon: a state of the art cybersecurity specific model, designed for vulnerability research, finding, prove, and fix vulnerabilities in real software. Poseidon preforms: - 17.6% higher than GPT-6 Astra - 2.9% higher than Claude Fable 5.2 (CWE-Bench)
44
48
6
658
57,059
Best way to destress during midterm season is to do aerobatic flights
2
65
The team blindsided me with a polygraph test, it was crazy. But it's important to me that Hark does this right. Your data is yours.
70
40
25
746
65,285
Two hands, one robot… or two? I built an oversimplified pipeline that turns a video demonstration into simulated robot actions, comparing one arm and two arm execution for the same video demonstration data. First test uses a Gemini generated video. Real world demos videos are next.
1
5
446
Teleoperate the rebot with ur mobile device with precision
4
1
22
1,211
Muse spark and download images, annotate and train on them and return me a yolo model…. And it’s not the best but it’s pretty decent for a one shot prompt
1
104
Getting my rebot to remove prints off my 3d printer bed no matter where I am in the world - autonomous mode - Teleoperate mode with just your iPhone or android device as the controller Will be open source in the next few days part by part
10
470
Just got the seed studio - drop ideas below of what I should get it to do first!!
4
117
Astra is great and all at coding but damn bro does it suck at following simple instructions
1
39
Flashing the connect tech rudi nx! Let’s see how it preforms! Yes it’s with jetpack 6.2.1 and yes ik it’s old 🤣
1
57
Had a great time at Roscon this past week! Here is a pic of me at the most memorable booth @dimensionalos !!
1
4
173
Demo day @viamrobotics making a robot play jenga allowing people all across the world to play with each other! And using @viamrobotics software and tools we accomplished this in ONLY 5 hours!!!!
1
11
1,556