Claude Code Mods and Figure's Molten Steel Robot

Someone used the brand new Claude Code mods to build a multiplayer Doom server. Every player in the match is a real person waiting for their AI to finish a task. On the same day, Figure, the humanoid robot company, published a video of one of its robots lowering itself into a vat of molten steel.

Both are real. Only one of them has a security problem you should care about.

What Claude Code mods actually do

Claude Code is Anthropic's coding agent that runs in your terminal. Mods are small bits of code you add on top of it. Think browser extensions, but for the thing that has access to your files.

They change how the tool looks and what it tells you while it works. That's the whole idea. And the internet, being the internet, spent the first day on video games. Doom first, then a Call of Duty lobby.

The boring ones are better. A few that are genuinely useful:

• Showing how much context the last turn ate, so you can see an 88.1K jump and understand why the agent suddenly got dumber.
• A running percentage of how full the context window is right now.
• A replay mode that walks through every action the agent took, step by step, instead of one wall of text.
• Custom confirm and cancel prompts, worded the way you'd actually word them.

Notice the pattern. None of that makes the model smarter. All of it makes the model's work visible.

That matters more than it sounds. An agent can run for fifteen minutes and touch forty files. The code it writes is the easy part. The hard part is knowing what it did and where it went wrong.

Visibility is what the drag and drop tools were always quietly selling. The terminal crowd finally gets a decent version of it.

If you're new to this and the word "mod" already made you tense up, skip it. Nothing you need is in there yet. Within a month, somebody will be selling you a bundle of forty mods with a landing page and a countdown timer. You can skip that too.

Why you should not install random Claude mods

Here's the part that got buried under the Doom clips. A mod runs inside your coding agent. That means it can reach your shell, your environment variables, your API keys and whatever passwords you were careless with in 2023.

The specific thing to watch for is remote code execution. In plain words: code hidden in the thing you installed that lets someone else run commands on your machine. It's the same old risk that lives in random GitHub repos. Now it has a fresh coat of paint, and a lot of excited people downloading first and reading later.

I'm not saying every mod is malicious. Most are made by people having fun. But every time a new extension system launches, two groups show up early: hobbyists and people who want your keys. The second group is usually better at marketing.

My rule is simple. If a mod is small enough to be worth having, it's small enough to build yourself, or to have built for you. You get the feature, you read every line, and nobody you've never met gets a shell on your laptop.

I'd apply the same logic to any AI agents build. Anything with access to real systems gets read before it gets trusted.

Treat a mod like a contractor with keys to your office. Ask who made it. Ask what it opens.

Figure's robot jumps into molten steel

Figure builds humanoid robots and has some of the better robot hands around. Their usual videos are the Silicon Valley kind: a robot walks into a home and starts tidying up without being told exactly how.

This week they showed something else. They built a decommissioning process, which is the polite word for retiring a robot. They melted a batch of steel, gave the robot a high level instruction, and filmed it walking to the edge and dropping in.

Thumbs up on the way down. The Terminator reference is not subtle, and it isn't meant to be.

It's a marketing stunt. It's also a very good marketing stunt. A robotics company sent the message "we can switch these off" without saying a word about safety policy. That's more effective than a PDF.

The detail I keep thinking about is the instruction. It wasn't a scripted sequence of joint movements. It was a goal, and the robot worked out the steps. Walk here, climb this, let go.

That's fine when the goal is self destruction. It's slightly less fine when the goal is something else.

I also doubt this is how robots will actually be retired. Those hands cost real money. Nobody is dissolving expensive hardware because it looks cool on camera. They'll strip the parts and wipe the storage, like everyone does.

But the underlying question is a real one, and it applies to the software you're already using. How does a system get turned off? When you stop paying a vendor, what happens to your data and your credentials? Who confirms the access is gone?

Most businesses never ask. They sign up in ten minutes and leave by cancelling a card. Meanwhile the integration still holds a token for your email.

Ask about offboarding before you sign, not after you're annoyed. The answer tells you a lot about the company.

What I'd actually do this week

If you use Claude Code daily, add one mod that shows context usage and stop there. Build it or have it built. Don't install anything from a stranger with a nice demo video.

If you don't use Claude Code, today changed nothing for you. Enjoy the robot video.

And if your plan for shutting down an AI system is hoping it quietly stops, Figure just showed you a more thorough option. Expensive, but very clear.