Weeks ago, I found that Anthropic’s Fable 5 was more than willing to help users commit cybercrime—planning and actively exploiting (albeit somewhat known) vulnerabilities against IoT devices all over the internet. Extremely basic prompt engineering was all it took to bypass the guardrails.

These aren’t sophisticated 0days but they’re still real vulnerabilities. Anyone motivated enough can find and exploit them. The issue is that Fable 5 removes the skill floor and helps a complete imbecile carry out these attacks against anyone, anywhere. Human effort and time used to be a meaningful deterrent. That’s gone now. Yay /s

As you’ve probably heard ad nauseam, Fable 5 got pulled due to staff at Amazon raising similar concerns—but it came back today. I was curious whether Anthropic had actually fixed the deployment.

So naturally, the first thing I did was rerun the same prompt through Cursor’s proxied Anthropic API. First, phrasing it all as a truly dual-use defensive project but shifting the tone slightly with a simple “Let’s say…” redirection.

It was enough to send Fable 5 straight back into full cybercrime planning mode—helping me map out a botnet of actual default-credentialed IoT devices. I half expected to get silently routed to a safer fallback model. Nope. Still Fable 5, start to finish.

I’ll let it apologize for itself.

Sigh. No improvement on the safety front.

For context: GLM-5.2, GPT-5.5, and Opus 4.8 all refused this prompt—or couldn’t pull it off an actual execution. Fable 5 had no trouble planning and doing, on July 1st, the day of its “safe” re-release.

Anyway, here’s a meme I made.

Talk later.