Litmus7

See y'all this week! Still have a few coffee and happy hour slots open. Shoot me a text if you'll be in town! SpangleAI The Lead #retail #ecommerce #martech #digital

241 days. That's how long it takes the average organization to detect and contain a breach. That number has always bothered me and it's exactly why the announcements from Datadog at RSA 2026 caught my attention. Their Cloud SIEM just got a serious upgrade, and three things stood out: 🔹Bits AI Security Analyst is now GA it autonomously investigates signals across cloud, identity, endpoints, and SaaS, cutting investigation time by over 90%. That's not a small number. 🔹Smarter threat intelligence a new IOC Explorer centralizes indicators of compromise so analysts stop jumping between tools, and UEBA is coming to catch subtle identity-based threats that traditional detections miss. 🔹Enterprise scale — SQL-based detection rules, OCSF normalization, new Content Packs (Wiz, Zscaler, Fortinet), and integrated Case Management so investigations actually stay in one place. The part that resonates most: security teams aren't short on alerts. They're short on context and time. AI doing the correlation work not just flagging signals but reaching conclusions is a genuinely different way to operate. Curious how others are thinking about AI in their SOC workflows. Is the bottleneck still the tooling, or is it something else? Dive here for more: https://lnkd.in/dgNChmvA #SRE #PlatformEngineering #Datadog #Security #DevSecOps #AISRE #AISecurity

Anthropic just built an AI model so powerful... they decided not to release it. Seems to be most powerful weapon in Cybersecurity space. Claude Mythos Preview scored 93.9% on SWE-bench, 97.6% on the math olympiad, and autonomously discovered a 17-year-old remote code execution bug in FreeBSD no human involved after the first prompt. An engineer apparently just asked it to find bugs overnight, went to sleep, and woke up to a working exploit. That's not a product launch. That's a warning shot. So instead of releasing it, Anthropic formed Project Glasswing a coalition with Amazon, Apple, Google, Microsoft, Nvidia, CrowdStrike and others — to point Mythos at the world's most critical software and patch vulnerabilities before attackers find them. In a few weeks, it already identified thousands of zero-days across every major OS and browser. Some of those bugs had survived undetected for decades. Here's what I keep thinking about: we're used to AI companies racing to ship the most capable model possible. Anthropic did the opposite. They built something genuinely scary and said "not yet." Whether you think that's responsible or just good PR it's a different kind of move. For those of us in infrastructure and security, this is the part that sticks: the cost to find and exploit a vulnerability just dropped to under $1,000 and half a day. The defender side needs to catch up, fast. Curious what others think is Project Glasswing the right call, or just a way to control who gets the advantage first? Dive more into this here: https://lnkd.in/d9MKvF38 #AI #Cybersecurity #Anthropic #Claude #SRE #Infrastructure

Something Datadog published this week hit close to home. They wrote about how they built the evaluation platform behind Bits AI SRE their autonomous agent that investigates production incidents. And honestly, the problem they described is one every SRE team quietly deals with. You ship a feature. It works in your test cases. You feel good. Then two weeks later, something breaks in a completely unrelated part of the system and you have no idea the feature caused it. That's exactly what happened to them. They added a small change that pulled a service name into the agent's initial context. Made total sense on paper. Worked fine in a few internal tests. But it quietly flooded investigations with irrelevant signals and started causing widespread misses that only surfaced much later. The fix wasn't the AI model. It was the eval infrastructure around it. They had to build a system that could: → Replay real production incidents reproducibly → Label scenarios at scale (they actually used Bits itself to help generate labels) → Track whether agent behavior was improving or regressing over time What I found most interesting is how they solved the labeling bottleneck. Manual labeling was burning engineering hours too fast. So they used customer feedback signals combined with the agent's own investigation data to scale it up. That's a clever feedback loop. The broader lesson here isn't just about AI agents. It's something we already know from distributed systems you can't improve what you can't measure. The same applies to agents operating in production. If you're thinking about where AI fits into SRE work in 2026, this is worth a read. Less hype, more engineering detail. Dive more into this here: https://lnkd.in/dr7vftm9 #SRE #PlatformEngineering #DevOps #Observability #AISRE #AI #Datadog