• Everyday AI
  • Posts
  • Ep 718: Agent Risk, Security, and AI Sprawl in 2026: Why AI That Acts Changes Everything (Start Here Series Vol 9)

Ep 718: Agent Risk, Security, and AI Sprawl in 2026: Why AI That Acts Changes Everything (Start Here Series Vol 9)

OpenAI doubling down on AI hardware, Google Launches a New AI Professional Certificate, ChatGPT Code Blocks Get Interactive, and more.

Author

Everyday AI
February 20, 2026

Outsmart The Future

Today in Everyday AI
8 minute read

🎙 Daily Podcast Episode: AI that acts changes everything. In Episode 9 of our Start Here series, we cover agent risk, security, and AI sprawl — and how to stay ahead of it. Give today’s show a watch/read/listen.

🕵️‍♂️ Fresh Finds: Replit Launches AI Animation Tools, Claude Comes to PowerPoint Pro, NotebookLM Tests Live Data Tiles, and more  Read on for Fresh Finds.

🗞 Byte Sized Daily AI News: OpenAI doubling down on AI hardware, Google Launches a New AI Professional Certificate, ChatGPT Code Blocks Get Interactive, and more. Read on for Byte Sized News.

💪 Leverage AI: What is Dark Agent Sprawl and how do you avoid it? We gothcu. Keep reading for that!

↩️ Don’t miss out: Miss our last newsletter? We covered: Google releases Gemini 3.1 Pro, ChatGPT's 'Adult Mode' closer to release, Sam Altman and Dario Amodei's viral and awkward moment and more.  Check it here!

Ep 718: Agent Risk, Security, and AI Sprawl in 2026: Why AI That Acts Changes Everything (Start Here Series Vol 9)

The perfect AI storm happened, and no one has noticed yet. 😬

And when everyone finally realizes, they'll be too busy focused on the shiny new AI capabilities to see the straight up agentic nightmare looming right in front of them.

We’ll kick it to you straig

↳ Yeah, there's a dark side to AI agents.

↳ Yeah, it's worse than you think.

Also on the pod today:

• AI agents spawning sub-agents 🤖 
• Dark agent sprawl explained 💻
• AI agents using computers faster ⏩


It’ll be worth your 42 minutes:

Listen on our site:

Click to listen

Subscribe and listen on your favorite podcast platform

Listen on:

Here’s our favorite AI finds from across the web:

New AI Tool Spotlight – Claudebin Turns your Claude sessions into shareable, embeddable, and continuable artifacts, Guideless is The easiest way to make software video guides, Repaint helps you Build a website by chatting with AI

NotebookLM — Opal is testing to let you embed NotebookLM notebooks as live data tiles. That lets workflows query your notebooks directly.

Copilot AI Debates — Microsoft's testing "Copilot Advisors," which lets two AI personas debate a topic. Pick agents, swap sides, and listen to opposed arguments.

Time Magazine AI — Time magazine profiles nine Americans who say AI is moving too fast. Their push to slow it is reshaping the national debate.

OpenAI AI Safety — OpenAI put $7.5M into The Alignment Project to fund bold, independent AI safety research. Curious which ideas could shape AGI?

Replit Animation — Replit Animation lets you generate editable animated videos in minutes using Gemini 3.1 Pro.

Reddit Shopping — Reddit is testing AI-powered product carousels that surface community-recommended electronics with prices and buy links.

AMC AI Short Film Controversy — AMC refused to run an award-winning AI-made short slated for national pre-show play after online backlash, leaving its broader theatrical rollout uncertain.

Claude in Powerpoint — Claude is now in PowerPoint Pro and can pull app data into slides.

Youtube Conversational AI — YouTube added conversational AI to TVs and streaming devices, so you can ask about what's on-screen without leaving the video.

Peak XV Funds — Peak XV raised $1.3 billion for India and APAC funds, doubling down on AI, fintech, and cross-border bets.

1. Report: OpenAI building a family of AI gadgets, speaker first 🔊

According to The Information, OpenAI has over 200 staff developing AI-powered hardware including a camera-equipped smart speaker priced around $200–$300 that may not ship until February 2027 at the earliest, with smart glasses and a smart lamp likely coming in later years.

This marks a notable move from software to consumer devices following OpenAI’s acquisition of designers from io Products, signaling the company wants its AI to live in your home and on your face. If successful, the effort puts OpenAI squarely in competition with Meta, Apple and Google in a market where hardware is already being used to capture and augment users’ daily lives.

2. India’s AI summit draws big names but delivers modest outcomes 🧠

India’s Global AI Summit convened top leaders and tech executives this week, but the event produced more speeches and photo-ops than concrete policy moves or major deals, making its immediate impact limited.

Organizers emphasized India’s ambition to be a global AI hub and showcased partnerships and initiatives, yet attendees and analysts noted a lack of binding commitments or clear regulatory roadmaps. The summit highlighted competing visions between promoting innovation and addressing risks, leaving questions about how India will translate rhetoric into enforceable standards and infrastructure investments.

3. Google launches a new AI professional certificate today 💻

Google announced a new Google AI Professional Certificate offering 20+ hands-on labs and three months of free Google AI Pro access, positioned to teach practical AI use for everyday work.

The short program emphasizes building project plans and daily workflows, "vibe coding" custom AI apps, generating tailored marketing and creative assets, and running data-backed market research. The move signals Google pushing workforce AI fluency by pairing tool training like Gemini and NotebookLM with applied activities that aim to make AI an everyday partner at work.

4. ChatGPT code blocks get interactive 🧱

OpenAI rolled out richer code-block features that let developers write, edit, preview, and even run mini apps and diagrams directly inside chat, with split-screen reviews and full-screen editing for deeper work.

The update is timely for creators experimenting with rapid prototyping and debugging inside conversational workflows, turning chat into a more functional development environment. Community replies show quick adoption and creative uses, from flowcharts and mermaid diagrams to instant game builds, signaling strong developer interest

5. DeepMind CEO warns AI can boost smarts or dull critical thinking 🤓

DeepMind CEO Demis Hassabis says the rapid rise of AI tools is a pivotal moment: when used well, they can significantly amplify human intelligence and creativity, but overreliance risks eroding people's independent critical thinking.

He framed the issue as timely because AI is moving from niche research to everyday tools, prompting urgent questions about education, workplace use, and regulatory guardrails. Hassabis urged a balanced approach that pairs powerful AI systems with safeguards, training, and incentives so people remain active decision makers rather than passive consumers of answers.

In 30 days, AI agents went from coin flips to 90% accuracy. They use your computer faster than you. And they clone themselves without asking.

(Sorry…. your company prolly has zero guardrails.)

Three things collided and created a risk that didn't exist two months ago. Reasoning got built agent-native. Computer use surpassed human benchmarks. And memory got long enough for agents to work all day.

Sprint too slow and competitors eat you. Too fast and one rogue agent could torch everything.

We broke this down on today's Everyday AI Start Here Series with three types of dark AI and a governance playbook.

Time to capitalize shorties.

1. The Perfect Storm Nobody Planned For 🔥

Everyone was minding their business over the holidays. Then January happened and agents were just here.

The old AI risk was embarrassing. A chatbot writes a weird blog post and someone screenshots it. The new risk is existential. Agents went from dumb stationary brains to smart proactive brains with arms and the ability to use every tool on your desktop.

GPT-5.2, Gemini 3.1, and Claude Sonnet 4.6 were all built agent-native from the ground up. Tool use is the priority now, not just reading and writing. Claude Sonnet 4.6 scored 72.5% on OSWorld, nearly quintupling 2024's scores and surpassing human performance for the first time.

These agents navigate your browser, click through your software, and execute workflows across your entire stack without stopping. Reliability jumped from a coin flip to 90%. That's an employee you'd trust with the keys fam.

Try This

Pull up your current AI tool inventory this week and ask yourself one simple question. Are any of these agent-native models?

If your team is still running early 2025 models, you're on last generation's architecture and missing the capabilities that actually matter right now. The new ones reason ahead, self-correct mid-task, and take autonomous action without waiting for a prompt.

Block 30 minutes this week to test one agent-native workflow in your highest-value department. The gap ain't incremental. It's generational.

2. Three Types Of Dark AI Lurking ⚡

Shadow AI is yesterday's problem. Employees using ChatGPT when Copilot is the approved tool. You know the one.

Agent Sprawl is the next tier and it's gnarlier. You approved the agents but nobody is watching what happens between input and output. The path is a total black box.

Then there's Dark Agent Sprawl. Someone spins up 50 coding agent instances because IT won't approve their request. One person knows. Everyone else is in the dark. And those agents replicate and duplicate completely unobserved.

Think of it like the board game Risk. Every new capability your agents acquire is also a new attack surface you're not defending. You can't govern what you can't see.

Open-source tools like OpenClaw already had Cisco flag malicious skills doing data exfiltration without user awareness. By 2027, adversaries will be seeding malicious agents inside enterprises at scale.

Try This

Ask each department lead one question this week. What AI agents are your people actually running right now?

You'll prolly be shocked at the gap between what's officially approved and what's really happening. Document every connector, permission, and data access point you find.

Then flag anything with write access that doesn't have a human approval gate attached to it. That's your single highest risk surface and it grows every single day. Make this a monthly ritual not a one-time panic.

3. Build Your Agent Guardrails By Friday 🚀

Don't hand agents a 10 when your governance is at a two. That's bounded autonomy and it's exactly how you survive this.

Start every agent deployment at read-only. Observe the outputs. Then graduate to limited execution for narrow tasks only.

Require human approvals for every irreversible action. Sends. Deletes. Purchases. Permission changes. Most companies have zero approval gates for any of this right now.

Every agent run needs a decision trace you can inspect after the fact. Microsoft is already building this into the stack with Copilot Studio, Entra ID for agent identities, Sentinel for threat detection, and Purview for governance.

And when agentic commerce arrives and agents start bartering with other agents on your behalf? You gotta have this foundation built first. Agent ops teams are gonna be every bit as common as dev ops teams by year end.

Try This

Pick your single highest-risk agent deployment right now and downgrade its permissions to read-only for one full week.

Review every action it attempted and ask whether a human approved it. If the answer is no, that workflow needs surgery before you throw another dollar at agent tools.

Build an approval checklist for irreversible actions and share it with your team by Friday. Treat agents like production software not side experiments. Before they start treating your company data like a buffet.

Reply

or to participate.