OpenAI backs AI cyber defense agents, FBI flags 400 million in crypto theft, and Gemini 4 Hydrogen lands
Compact Conversations for 2026-11-12: 5 AI stories, ai news worth knowing in just 5 minutes.
[Audio embed placeholder]
The Lead: Cogent and OpenAI build cyber defense factories for enterprises
Cogent Security uses OpenAI models to build Cyber Defense Factories that autonomously find and fix vulnerabilities. The system maintains a knowledge graph per customer and uses GPT-6.1 Sol for execution and GPT-5.6 for routine triage. Cogent reports a Fortune 1000 automotive customer reduced mean time to remediate critical vulnerabilities by 75% and a Fortune 500 IT services customer achieved 5x vulnerability-triage throughput.
Why it matters: Cogent’s approach shifts enterprise security from finding vulnerabilities to fixing them at machine speed, with AI agents handling investigation through verified remediation.
Source: OpenAI Announcements
The Feed
FBI says North Korean hackers stole $400M+ in crypto using deepfake video calls
The FBI says North Korean state-backed hackers stole over $400 million in cryptocurrency over 18 months using a scheme called FaceStealer. Targets are invited to video calls from compromised contacts on fake meeting software, then tricked into clicking a link that infects their machine. The group, Sapphire Sleet, is now using AI deepfakes of crypto industry figures.
Why it matters: The FBI warns that even video calls from known contacts can be weaponized, and AI deepfakes are making social engineering harder to detect.
Source: Reuters
DeepMind launched AlphaGenome 2 for DNA analysis cell by cell
Google DeepMind released AlphaGenome 2, an AI model that predicts what a single-letter DNA change does in each cell type. It learned from over 100 million individual cell profiles across 68 tissues. DeepMind reports it found nearly twice as many gene-changing variants as the previous best model. The model is for research use only, not clinical or diagnostic work.
Why it matters: AlphaGenome 2 lets scientists ask which cells a DNA variant affects, not just whether it’s harmful, which could improve target selection for drug development.
Source: The Neuron
Agility Robotics goes public as first US specialist humanoid maker
Agility Robotics became the first publicly listed humanoid robotics company in the US, trading on Nasdaq under AGLT. The offering raised about $465 million. The company has over $300 million in conditional multiyear orders for its Digit 5 robot, with about 100 units deployed in commercial environments.
Why it matters: Agility’s IPO signals growing investor confidence in humanoid robots for industrial applications, with a focus on manufacturing and logistics rather than domestic use.
Source: aibusiness
Google introduces Gemini 4 Hydrogen for fast, low-cost AI
Google launched Gemini 4 Hydrogen, a fast and efficient model priced at $1.50 per million input tokens across a 1 million-token context window. It surpasses Argon on the Artificial Analysis Intelligence Index at default settings and is designed for everyday coding and agentic workflows. Hydrogen is available across Google AI Studio, the Gemini API, Gemini for business, and Search.
Why it matters: Hydrogen provides a lower-cost, faster alternative to frontier models for routine tasks, and Google’s Orchestrator can automatically switch between models to optimize cost and speed.
Source: Google Blog
One Thing to Try
Microsoft is removing older local agent integration points from VS Code in early 2027, moving everything to the newer Copilot model. Developers on forums are frustrated by the pace. Take 15 minutes to check whether any extension, script, or custom agent you rely on depends on an integration point with a published removal date. If it does, test it against the replacement path now to avoid disruption.
Sources
- Cogent and OpenAI build cyber defense factories for enterprises - OpenAI Announcements
- The FBI says North Korean hackers stole $400M+ in crypto in the last 18 months using “FaceStealer”, a scheme involving fake video calls and - Reuters
- 😺 DeepMind launched AlphaGenome 2 for DNA - The Neuron
- Agility Robotics goes public as first U.S. specialist humanoid maker - aibusiness
- Gemini 4 Hydrogen - Google
- Local harness removed in Feb. 2027 - visualstudio.com
Transcript
Host A: Welcome to Compact Conversations, the show that compresses the day’s AI news into 5 minutes.
Host A: [conversational] Today’s lead comes from OpenAI’s own announcements: a startup called Cogent Security is building what it calls Cyber Defense Factories for enterprises, and it runs them on OpenAI models. The idea is to let AI agents handle security work end to end, from investigation through verified remediation. Cogent’s argument is that the hard part of enterprise security isn’t finding vulnerabilities anymore, it’s fixing them, because remediation spans teams and needs real context about which systems matter. Host B: [curious] And that context is where the product design gets interesting. Cogent keeps a knowledge graph per customer covering assets, vulnerabilities, identities, and dependencies, with background agents keeping it current across security tools. Its agents use OpenAI’s GPT 6.1 Sol for execution and a smaller model for routine triage. Every change gets checked against customer-defined policies, and higher-risk work can run in isolated sandboxes. Cogent reports a Fortune 1000 automotive company cut its mean time to remediate critical vulnerabilities by seventy-five percent.
Host B: [with a small lift] One number to know today, from that same lead story: five times. That’s how much more vulnerability triage throughput a Fortune 500 IT services customer achieved with Cogent’s Cyber Defense Factories, according to Cogent. Triage throughput is how many vulnerabilities the team can assess and prioritize in a given period. It’s Cogent’s own reported figure, not an independent audit, but it shows the kind of scale AI agents can bring to security operations.
Host A: [thoughtful] Other stories from the feed, starting with Reuters reporting on an FBI advisory out today. The Bureau says North Korean state-backed hackers stole more than four hundred million dollars in cryptocurrency over the past eighteen months using a scheme it calls FaceStealer. Targets get invited to video calls from real contacts whose accounts were compromised, but on imposter versions of meeting software. Once on the call, a fake technical error prompts the victim to click a fix link, infecting their machine and social accounts. The FBI says the group, known as Sapphire Sleet, is now using AI deepfakes of well-known crypto figures. Host B: [quick aside] The FBI’s practical warning is simple: don’t trust a video call just because the invitation came from a real contact. Host A: [curious] Next up, Google DeepMind released AlphaGenome 2, an AI model for predicting what a single-letter DNA change actually does, cell type by cell type. It learned from over one hundred million individual cell profiles across sixty-eight tissues in humans and mice. DeepMind reports it found nearly twice as many gene-changing variants as the previous best model, and it’s for research use only, not clinical or diagnostic work. Host B: [lighter] In robotics news, Agility Robotics has become the first publicly listed humanoid robotics company in the US, trading on Nasdaq. It raised about four hundred sixty-five million dollars and says it has more than three hundred million in conditional multiyear orders for its Digit 5 robot. Host A: And our last item: Google introduced Gemini 4 Hydrogen, a fast, lower-cost companion to its Carbon and Argon frontier models, priced at a dollar fifty per million input tokens across a one-million-token context window. It’s being introduced across the Gemini app, the API, and Search.
Host B: [conversational] One Thing to Try, and this one is aimed at anyone who builds on tools that change underneath them. Microsoft is removing the older local agent integration points from VS Code in early 2027, moving everything to the newer Copilot model, and developers on the forums are not thrilled about the pace. So here’s the habit: take fifteen minutes and check whether anything you rely on, an extension, a script, a custom agent, depends on an integration point with a published removal date. If it does, run it against the replacement path now.
Host A: That’s Compact Conversations for Thursday. More AI news tomorrow. Until then, happy prompting.