28 August 2026

A Nameless AI Model and the Inbox I Tried to Hack

A free coding model appeared anonymously for six days before the community fingerprinted it as GLM-5.3-Flash — and I spent a week trying to break into my own AI email agent.

---

Dear Reader,

Two posts this week, and the loose thread between them is trust: how much of it you should extend to a system before you know who built it or what it will do with the access you’ve given it.

One story is about a model that wouldn’t say its name. The other is about an agent I built myself and then tried to trick.

---

AI Weekly Signals: The Free Model Nobody Would Name https://daehnhardt.com/blog/2026/08/28/ai-weekly-signals-free-model-nobody-would-name/

A free, unbranded coding model called Ox Alpha showed up on OpenRouter offering near-unlimited rate limits, the kind of deal that should make anyone suspicious. It took a developer’s fingerprinting tool and six days of unpaid community testing before Z.ai revealed it was GLM-5.3-Flash. In the same week, the same Claude Opus 5 that scores about 30% on ARC-AGI-3 reached 100% inside NVIDIA’s AVO harness, Anthropic let its most capable security model loose behind a scan-only interface, and a 23-year-old’s AI assistant raised $250 million days after testers found it still reading their email once they’d disconnected it.

I Gave My AI Agent the Keys to My Inbox — Then Tried to Break Into It Myself https://daehnhardt.com/blog/2026/08/24/i-gave-my-ai-agent-the-keys-to-my-inbox/

If an unidentified model asking for trust made you uneasy, this is the other side of that coin: what happens once you’ve already handed an AI agent the keys. I built an email agent, then set out to red-team it myself with prompt-injection attacks hidden inside test emails, the same trick that worried me most about Instinct’s assistant in the Signals post above. What I learned is that securing an agent is system architecture, not better prompting — and I walk through the guard I built to catch and log the attempts.

---

Do hit reply: would you use an AI model if you didn’t know who operated it, provided the performance was good enough?

Have a great weekend, and thank you very much for reading,

Elena

---

← All issues