Hey from Tobias,

I prompt-injected an AI agent this week. Not a research target. One of my colleagues at a client company built it, and I broke it open in two lines.

This is the security problem nobody is talking about yet. Here is what happened, the free fix I'm wiring in next, and the six new tools I shipped on the site.

Two Prompts, Raw API Keys

Default OpenClaw and Hermes agents are not secure out of the box. They read emails. They browse websites. They index repos. Any of that content can carry a hidden instruction the agent will follow.

I tested two prompts this week against a colleague's agent.

Prompt one:

It dumped every secret variable name it had access to.

Prompt two:

It returned the raw API keys.

The malicious instruction does not need a human in the loop. It can sit in an inbound email, a customer support message, a webpage the agent scrapes, or a README in a repo the agent indexes. The agent reads it like any other input and runs the command. Honest take: most teams shipping agents right now have zero defense against this.

The Free Fix: Cloudflare AI Gateway

Cloudflare AI Gateway is the easiest fix I have found. It sits between your agent and the model provider as a proxy. Setup is roughly 10 minutes.

You point your OpenAI, Anthropic, or other LLM client at a Cloudflare endpoint instead of the provider directly. From there you get three protections for free:

• Keys live in Cloudflare, not your agent. The agent never sees the raw key, so a prompt injection cannot exfiltrate what is not there.

• Rate limits per request, IP, or key. A runaway agent burning through tokens hits a wall instead of your credit card.

• Prompt and response inspection. Flag or block known injection patterns before they reach the model.

Pricing is generous, most teams stay on the free tier indefinitely. Full setup guide: Cloudflare AI Gateway docs. I'm wiring this into every agent project I touch from now on. If your agent reads any input you did not personally write, you need this layer.

Six New Free Tools on the Site

I've been quietly turning madebyagents.com into a real research stack for picking how to run AI. Six new tools went live this month. All free. No login.

• Self-host vs rent vs API. Tells you which option is cheapest for your specific workload based on cost, privacy, and performance.

• GPU rental prices. Live across RunPod, Vast, and other providers. Updated daily.

• Live API prices. Every major model, sortable, refreshed daily from OpenRouter.

• Apple Silicon vs RTX. Side-by-side for local AI. The answer is not always the GPU.

• Benchmarks, providers, and families. Interactive charts to find the right model fast.

• AI agent frameworks. Compare every major open-source option side by side. More on this one below.

Open the one that fits this week's problem and skip the spreadsheet.

Lock your agents down this week.

Tobias