Videos
Hard truths,
about a minute each
The failures I see in platforms, Kubernetes and AI agents in production. Each video covers one problem and what I would check first. Hover or tap to play; click again for sound.
Walkthroughs
Can an AI agent delete a protected AWS table?
I turned on deletion protection. Claude Code, an AI coding agent, switched it off and deleted all three test tables: the test user I gave it was allowed to change settings. Then I tightened its permissions twice. Both times the protected table survived.
Claude vs Codex vs Gemma: who fixes Kubernetes fastest?
I broke one setting in my Kubernetes app, then timed three AI agents fixing it on my Mac: Claude Code, Codex and a local Gemma, 18 runs in all. Every run fixed it. Claude Code was quickest, 13 seconds in a typical run; Gemma with thinking on took 1 minute 26. My experiment, not a benchmark.
Bedrock Guardrails masked the email. Is it still in your logs?
I tested Amazon Bedrock Guardrails on a made-up customer email. It found the address and masked it. The catch: the trace keeps the original, so if your app logs the whole result, the email ends up in your logs.
Can an AI agent find the fault?
I broke a demo Kubernetes cluster three ways. The agent got the built-in view role: it reads most things, changes nothing, and it can’t open secrets. It found three out of three. I picked the faults, so it’s no benchmark.
Same Kubernetes fix, nearly 3x the price.
Claude Code fixed it in 13 seconds for 6.7¢. Codex took 18 seconds and 18.9¢.
Has your website got the PHP fix?
It’s not your site’s code, it’s PHP itself. The fix is out: ask whoever hosts your site.
AI deleted my protected table.
Deletion protection was on. The agent switched it off and deleted all three tables.
Email masked. Still in your logs?
The reply hides the email. A log written before the check still has it.
I switched on thinking. It took almost 3x longer.
Same fix by the AI on my Mac: 31 seconds thinking off, 1:26 thinking on.
Can an AI agent find a Kubernetes fault?
I broke a demo cluster three ways. Read-only, it found all three.
Is a sandboxed AI agent actually safe?
Limited to one folder, it can still be tricked into sending a file out.
Why does your AI quote the old price?
You updated the price list. Your AI might read an old copy.
Why not send an update to everyone at once?
It passed its tests. Real customers take paths nobody expected.
One metric label can create a huge pile of data.
One ID per AI chat sounds useful. Each new ID is another thing to track.
Your backup AI can give wrong advice.
The app stays online. The backup was only tested on simple questions.
Does your AI only pass rehearsed tests?
The AI passed all 20 questions. Was yours one of them?
Slow to adopt AI, or just being careful?
Teams slow on AI aren’t laggards. They’re careful.
Does disconnecting an app make your AI forget?
Disconnect the app. Your AI may keep what it read.
What does an AI agent actually do?
The AI suggests what to look for. The software around it runs the search.
Did AI actually look that up?
A convincing answer can still be wrong. Ask where it came from.
Why does AI give two different plans?
You ask the same holiday plan twice, but get two different trips.
Why does AI give such generic answers?
You ask a vague question, then expect the perfect answer.
Why are you still fixing what AI wrote?
Give it a clear job and some boundaries, then check before you send.
Is AI actually saving you time?
AI took 10 seconds. How long did you spend?
Would you let AI build your weekly shop?
AI helps build your food shop. You still place the order.
Your health check can make an outage worse.
A slow database, a failing check, and a restart loop.
“Just get a human to review it” doesn’t scale.
AI writes changes faster than anyone can really read them.
Your app is slow and the CPU is idle.
A container CPU limit can make it wait.
Finishing the task isn’t understanding it.
An agent can follow every step and still miss what you meant.
Private endpoint won’t connect? Check DNS first.
Before you touch a firewall rule.
One retry, two tickets.
A timeout doesn’t necessarily mean the action failed.
Your AI agent can take orders from a web page.
Indirect prompt injection, and how to contain it.
Does stop actually stop your AI agent?
The chat goes quiet. The background job may not.
Your observability is lying to you.
Every pod green, and nobody can log in.
An accurate AI answer can still be a breach.
The service account could open it. The person asking couldn't.
Agent demos hide what happens on retry.
The connection drops mid-task. Hit retry: one ticket or two?
Your AI agent needs a way to stop.
A token limit controls one call, not the whole run.
A healthy agent can still do the wrong thing.
Monitoring shows it is up, not that it chose well.
Where does your AI agent actually run?
The model and its tools don’t always live in the same place.