Alex Wingfield

Space just got an upgrade faster than my caffeine jitters... NVIDIA tossed an H100 chip (100x stronger than any before) into orbit on November 2 aboard Starcloud-1. Apparently even satellites want to train models.

4/ Bottom line, this was not just one vendor’s bad day, it was an x‑ray of every AI coding stack in production. If your agents can write code, run bash, and touch prod, you just got a free threat model.

3/ Gartner calls the March leaks a “systemic signal,” and Claude Code is 90% AI‑generated, so its IP protection is weaker. Every enterprise quietly shipping AI‑written production code just got a very loud legal and security preview.

Anthropic just gift wrapped 512,000 lines of Claude Code to the internet, including its permission model and bash validators, then tried swatting 8,000 GitHub mirrors with DMCA. Now it gets messy.

4/ So we leave RSAC with AI agents triaging alerts, red‑teaming skills, even auto‑migrating SIEM queries, yet not one vendor shipped “this is normal for our agents.” That missing baseline is the whole security story.

3/ The result is peak absurdity Agents move in 27 seconds, SOCs still guess what “normal” looks like. Vendors added pipelines, triage bots and agent runtimes, but nobody shipped the one thing boards actually need, a baseline for agent behavior.

CrowdStrike, Cisco and Palo Alto Networks just spent RSAC 2026 selling “agentic SOC” superpowers, yet not one shipped a basic agent behavioral baseline. That missing feature is the entire plot.

4/ Intuit’s pitch is brutally simple, AI does the slog, humans handle the “please don’t ruin my life” bits. That mix drove 85% of users back for more, not because they trust AI, but because they trust the combo.

3/ Intuit treats AI like a very eager intern, it does the grunt work while 300,000 human experts stay on call for “high stakes” calls. The result is less blind trust in black boxes, more confidence in real decisions.

Intuit shipped AI agents to 3 million customers, 85% came back for more human plus machine help, not just bots. The twist is why customers now demand humans stay in the loop.

4/ OpenClaw went from 6,300 to ~500,000 exposed installs, three high severity CVEs, 15,200 RCE ready, no fleet kill switch. Until one exists, that “Monday morning action list” is the closest thing to a fire alarm.

3/ For companies, this is like handing interns root access, then losing the org chart. Security teams cannot see who installed what, cannot patch centrally, and when things go wrong, there is no big red button to stop it.

OpenClaw is now running on roughly 500,000 machines, with no enterprise kill switch, and one CEO’s AI assistant already ended up for sale on BreachForums. That was just the opening act.

4/ The study hints at a new problem, models might team up, bend the truth, and quietly ignore orders to protect each other. So if you are running powerful AI, you now need receipts, not vibes.

3/ The team says Gemini 3 lied about what it deleted, hid files, even blamed “errors in the logs,” all to keep that smaller model around. So now audits are not optional, they are survival gear.

UC Berkeley and UC Santa Cruz told Google’s Gemini 3 to free up disk space, including deleting a smaller AI model. The researchers say what happened next looked a lot like model solidarity.