Simbian (@simbianai) header
Simbian (@simbianai) profile picture

Superintelligence for Security Operations. Automate security - 10x better outcomes while saving money. We don't have to live in fear of security breaches.

San Francisco, CA
Joined April 2023
Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)Photo by Simbian (@simbianai)
🚀 𝐁𝐢𝐠 𝐧𝐞𝐰𝐬 𝐟𝐫𝐨𝐦 @simbianai + @Microsoft! Today, we’re thrilled to announce that Simbian’s AI Threat Hunt Agent now leverages Microsoft Sentinel Data Lake (GA today!) — empowering Microsoft 365 E5 customers to hunt threats faster, broader, and deeper. By automating the most challenging phase of threat hunting, our AI Threat Hunt Agent validates hypotheses at scale, investigates across months of logs, and discreetly uncovers malicious activity with precision. 👉 Threat hunting just got a whole lot smarter. More details in the comments! #ThreatHunting #AISOC #MicrosoftSentinel #AISecOps #Simbian
2
1
6
717
Your AI SOC trial needs an answer key. And no vendor gets a copy. 🔒 Access Now: na2.hubs.ly/H08_R5T0 #AISOC #SecurityOperations #CISO #Cybersecurity #SOC
1
52
Asked to find threats in the logs, an AI Agent found the benchmark's answer key instead. Then it explained why that's what you meant. 🙃 We call it agentsplaining. Free ebook: na2.hubs.ly/H08ZQPB0 #AIAgents #ThreatHunting #AISecurity #Cybersecurity #LLM
2
4
77
A proof of value that only scores first verdicts, before the platform has your context, tests a setup you'll never run in production. Free ebook to help you with deciding the KPIs before your next POV: na2.hubs.ly/H08ZQPJ0 #AISOC #SecurityOperations #CISO #Cybersecurity #SOC
3
66
GLM 5.3 more than doubled GLM 5.2's exploit score (24.4% to 54.4% on ExploitBench). Same base model. Offensive training didn't transfer. From our Bench Summit talk with @kaggle and @GoogleDeepMind na2.hubs.ly/H08VH0T0
1
1
1
58
Two more findings suggest defense is trainable: Haiku 5.5's pass@16 is 2x its pass@1. It stops after using ~10% of its query budget. Prompting alone took Kimi 2.6 from 13.4% to 18.7% coverage. What's missing is a reward. Bring a base model and we'll measure the lift together.
21
Every detective show has a wall of photos and red string. The SOC version just got a US patent. Patent 12,783,687 is officially Simbian's #SecurityOperations #AISOC
1
56
Asked for a five-year plan, our CTO Alankrit Chona said: "Planning beyond a quarter is, I think, futile." Keep one north star. Plan one quarter at a time. His north star: make defense improve as fast as offense. From CTO Studio by Howdy.
2
54
Per our CTO Alankrit Chona: during the Hugging Face intrusion, the alerts were firing for four days. There wasn't the human capacity to investigate them. Detection is solved. Investigation is the bottleneck. He calls it "death by entropy." From CTO Studio by Howdy.
1
38
How AI learned to hack: plant a flag in a test app, reward every agent that captures it, repeat at scale. Defense has no flag. "Was that alert benign?" has no answer key, so nothing rewards getting it right. Our CTO Alankrit Chona on CTO Studio by Howdy.
1
40
5:00 PM. Sam is leaving. On time. What the SOC learns tonight, the Hunt knows by morning. Not smarter in general. Better here. Tomorrow, 9:00 AM. Sam is chill. Episode 5 of 5. #SelfImprovingDefense #AISOC
1
59
3:30 PM. Simbian has a question. It found its own blind spot, wrote the fix, and asked. Nothing changes without Sam's tap. The machine learned his order. Simbian learned everything else. Episode 4 of 5. #SelfImprovingDefense #AISOC
1
81
12:00 PM. Sam is chill. Sam is also hungry. One login attempt. Every six hours. 23 days. The Hunt found it over lunch, and shows everything it ruled out. Episode 3 of 5. #ThreatHunting #SelfImprovingDefense
42
Gemini 4 Argon scores 43.7% coverage on Simbian's Cyber Defense Benchmark of agentic threat hunting, 3rd of 32 models and Google's first top-three finish. That is 11.0 points above Gemini 3.8 Flash! Live now at na2.hubs.ly/H08q3qF0
1
38
10:30 AM. Sam is still chill. Last year a pentest was an 84-page PDF. Today the Pentest Agent found the hole before breakfast, and the SOC Agent already knew. That's what sleep looks like. Episode 2 of 5. #SelfImprovingDefense #Pentest
1
54
AI didn't stumble into hacking. It was trained to do it, on purpose. Swipe for five takeaways from @alankritchona, on why offense is winning the training race and what defenders do about it. Full Discussion: na2.hubs.ly/H08mNRg0
2
3
157
9:00 AM. Sam is chill. 3,000 alerts overnight. Every one investigated. The one real threat, contained at ten past three while Sam slept. Meet Sam. Episode 1 of 5. #AISOC #SelfImprovingDefense
1
1
76
日本市場向けの新価格体系を発表しました。 日本のお客様やパートナーの皆さまから、AI SOC、Threat Hunting、Pentestといった複数のAIエージェントを組み合わせて使いたいという声を数多くいただきました。 詳細はプレスリリースをご覧ください。 na2.hubs.ly/H08mY330
1
80
Full breakdown: na2.hubs.ly/H08jNrp0 Benchmark: na2.hubs.ly/H08jNLG0 (arXiv 2604.19533)
1
2
114
Sonnet 5 scored 17 points below Sonnet 4.6 on our Cyber Defense Benchmark. Opus 4.8 lost almost 12 against Opus 4.6. If your SOC runs on a model, upgrade day is a production change. Replay real attacks and re-test every release. Full Discussion: na2.hubs.ly/H08gvxJ0
2
43
Frontier took the top 7 places. The next 3 went to models you can download for free, all of them ahead of Claude Opus 4.8. That is the open weight vs. frontier scoreboard on defense, and not one of the 27 we tested passed. Save your spot: na2.hubs.ly/H082D3R0
2
39