LessWrong posts by zvi

“Claude 4 You: Safety and Alignment” by Zvi


Listen Later

Unlike everyone else, Anthropic actually Does (Some of) the Research. That means they report all the insane behaviors you can potentially get their models to do, what causes those behaviors, how they addressed this and what we can learn. It is a treasure trove. And then they react reasonably, in this case imposing their ASL-3 safeguards on Opus 4. That's right, Opus. We are so back.

Yes, there are some rather troubling behaviors that Opus can do if given the proper provocations. If you tell it to ‘take initiative,’ hook it up to various tools, and then tell it to fabricate the data for a pharmaceutical study or build a bioweapon or what not, or fooling Opus into thinking that's what you are doing, it might alert the authorities or try to cut off your access. And That's Terrible, completely not intended behavior, we agree it shouldn’t do [...]

---

Outline:

(03:44) Introducing Claude 4 Opus and Claude 4 Sonnet

(06:08) Activate Safety Level Three

(08:17) The Spirit of the RSP

(09:46) An Abundance of Caution

(10:34) Okay What Are These ASL-3 Precautions

(16:55) How Annoying Will This ASL-3 Business Be In Practice?

(18:34) Overview Of The Safety Testing Process

(20:12) False Negatives On Single-Turn Requests

(20:39) False Positives on Single-Turn Requests

(21:33) Ambiguous Requests and Multi-Turn Testing.

(22:21) Child Safety

(22:55) Political Sycophancy and Discrimination

(25:46) Agentic Safety Against Misuse

(26:54) Alignment

(27:40) The Clearly Good News

(30:40) Reasoning Faithfulness Remains Unchanged

(31:47) Self-Preservation Attempts

(44:12) High Agency Behavior

(01:10:37) Oh Now You Demand Labs Take Responsibility For Their Models

(01:18:32) In The Beginning The Universe Was Created, This Made a Lot Of People Very Angry And Has Been Widely Regarded as a Bad Move

(01:30:52) Erratic Behavior and Stated Goals in Testing

(01:36:43) Situational Awareness

(01:37:11) Insufficiently Mostly Harmless Due To Then-Omitted Data

(01:40:15) Apollo Evaluation

(01:43:36) Model Welfare

(01:48:50) The RSP Evaluations and ASL Classifications

(01:55:22) Pobody's Nerfect

(01:56:54) Danger, And That's Good Actually

---

First published:

May 25th, 2025

Source:

https://www.lesswrong.com/posts/PjeZxCivuoyKhs4JB/claude-4-you-safety-and-alignment

---

Narrated by TYPE III AUDIO.

---

Images from the article:

Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts, or another podcast app.

...more
View all episodesView all episodes
Download on the App Store

LessWrong posts by zviBy zvi