In January 2024, I warned that the politics and inadequacy of guardrails would become a central issue for our times.

Gary Marcus@GaryMarcus

Five trends that will occupy GenAI in 2024: • The politics & inadequacy of guardrails • Copyright litigation • Customer retention • Lack of a moat as major players converge on roughly similar performance • Deepfakes (elections, fraud & nonconsensual porn)

2:40 PM · Feb 28, 2024 · 14.8K Views

---

16 Replies · 18 Reposts · 108 Likes

It took longer than I expected, but well, here we are:

WIRED@WIRED

Trump administration officials tell WIRED that if Anthropic wants to rerelease Fable 5, it will need to ensure the model's guardrails can't be circumvented. Security experts say that can't be done.

wired.com

The White House Wants Anthropic to Block All Jailbreaks. That May Not Be Possible

5:05 PM · Jun 17, 2026 · 4.97K Views

---

4 Replies · 3 Reposts · 15 Likes

Where do we go from here? At least with respect to LLMs, the security experts are right. And the writing has been on the wall literally for years. As Katie Conrad and I wrote here in January 2024:

virtually any guardrail has to thread a needle between the Scylla of being too restrictive and Charybdis of being too permissive. None thus far have done this effectively.

That’s still true. Next-token predictors simply aren‘t built for safety.

Either we curtail LLMs until we find a better technology, or we live with consequences.

Importantly, this is not an Anthropic problem, it’s a Generative AI problem.