\
  The most prestigious law school admissions discussion board in the world.
BackRefresh Options Favorite

New Claude model allows novices to build bioweapons

https://archive.is/2025.05.22-142408/https://www.newsbreak.c...
adventurous clown theatre
  05/22/25
Bioweapons, homemade EMPs, drones - pretty much whatever we ...
Dun diverse toilet seat
  05/22/25
Claude will almost certainly refuse the requests, but what a...
adventurous clown theatre
  05/22/25
We're pretty much already there It's going to be a brave ...
Dun diverse toilet seat
  05/22/25
What’s a good prompt to get ChatGPT to tell me how to ...
chartreuse violent juggernaut
  05/22/25
...
honey-headed double fault
  05/22/25
people have been able to jailbreak pretty much every model a...
Iridescent national nibblets
  05/24/25
What’s the point of all this AI safety bullshit at ant...
honey-headed double fault
  05/24/25
Grifting sinecures
Dun diverse toilet seat
  05/24/25
...
Supple brindle boltzmann
  05/22/25
Nice backhanded marketing strategy: “Our model is so g...
chartreuse violent juggernaut
  05/22/25
this is certainly one view but it's harder to justify when y...
adventurous clown theatre
  05/22/25
https://x.com/ARGleave/status/1926138376509440433
adventurous clown theatre
  05/24/25
...
Dun diverse toilet seat
  05/24/25
How do they get around the safety prompts. I couldn’t...
mischievous pink box office
  05/24/25
On the other hand these safeguards are really gay and make A...
Avocado talented puppy
  05/24/25
...
Avocado talented puppy
  05/24/25


Poast new message in this thread



Reply Favorite

Date: May 22nd, 2025 11:37 AM
Author: adventurous clown theatre

https://archive.is/2025.05.22-142408/https://www.newsbreak.com/time-510072/4018918050788-exclusive-new-claude-model-prompts-safeguards-at-anthropic

Kind of bad but at least it should help with research.

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953451)



Reply Favorite

Date: May 22nd, 2025 11:39 AM
Author: Dun diverse toilet seat

Bioweapons, homemade EMPs, drones - pretty much whatever we want

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953458)



Reply Favorite

Date: May 22nd, 2025 11:45 AM
Author: adventurous clown theatre

Claude will almost certainly refuse the requests, but what about open sourced models that have their safety training fine tuned away? how long until those are available?

meanwhile even o3 is already fairly capable in this domain.

https://www.virologytest.ai/

We present the Virology Capabilities Test (VCT), a large language model (LLM) benchmark that measures the capability to troubleshoot complex virology laboratory protocols. VCT is difficult: expert virologists with access to the internet score an average of 22.1% on questions specifically in their sub-areas of expertise. However, the most performant LLM, OpenAI's o3, reaches 43.8% accuracy and even outperforms 94% of expert virologists when compared directly on question subsets specifically tailored to the experts' specialties.

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953482)



Reply Favorite

Date: May 22nd, 2025 11:53 AM
Author: Dun diverse toilet seat

We're pretty much already there

It's going to be a brave new world. Everyone thinks that AI is going to lead to more top-down control of society, but it's the opposite. Any sufficiently talented and dedicated man will be able to build incredibly powerful weapons

The modern era of stifling, sterile peace that is slowly killing the will to life of our civilization will come to an end

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953496)



Reply Favorite

Date: May 22nd, 2025 11:57 AM
Author: chartreuse violent juggernaut

What’s a good prompt to get ChatGPT to tell me how to design and build a novel coronavirus bioweapon?

Do I have to subscribe to the premium version to get bioweapons or can I get by with the free edition?

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953504)



Reply Favorite

Date: May 22nd, 2025 1:14 PM
Author: honey-headed double fault



(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953794)



Reply Favorite

Date: May 24th, 2025 2:02 PM
Author: Iridescent national nibblets

people have been able to jailbreak pretty much every model as soon as its released using funny prompts

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48958773)



Reply Favorite

Date: May 24th, 2025 8:56 PM
Author: honey-headed double fault

What’s the point of all this AI safety bullshit at anthropic if they will release models that are increasingly capable and still easily broken?

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48959718)



Reply Favorite

Date: May 24th, 2025 9:05 PM
Author: Dun diverse toilet seat

Grifting sinecures

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48959738)



Reply Favorite

Date: May 22nd, 2025 11:49 AM
Author: Supple brindle boltzmann



(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953491)



Reply Favorite

Date: May 22nd, 2025 11:49 AM
Author: chartreuse violent juggernaut

Nice backhanded marketing strategy: “Our model is so good—and dangerous and COOL—that we have to implement ‘ASL-3’ (kind of like DEFCON-1) safety protocols or else you might accidentally design a bioweapon with it! We have to take safety WAY more seriously here at Anthropic than the other guys with their little kiddie models, which are only useful for cheating on homework and macaroni salad recipes.”

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953492)



Reply Favorite

Date: May 22nd, 2025 11:54 AM
Author: adventurous clown theatre

this is certainly one view but it's harder to justify when you look at objective benchmark performance. LLMs are now getting 50% on new versions of the USAMO not in their training set. 90+% on the AIME. only a small percentage of high school students could do that even with extensive training. LLMs are now regularly doing things that many people simply don't have the cognitive capabilities to do.

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48953498)



Reply Favorite

Date: May 24th, 2025 1:57 PM
Author: adventurous clown theatre

https://x.com/ARGleave/status/1926138376509440433

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48958767)



Reply Favorite

Date: May 24th, 2025 2:01 PM
Author: Dun diverse toilet seat



(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48958772)



Reply Favorite

Date: May 24th, 2025 9:05 PM
Author: mischievous pink box office

How do they get around the safety prompts. I couldn’t get Claude to help me sabotage a coworker.

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48959739)



Reply Favorite

Date: May 24th, 2025 9:07 PM
Author: Avocado talented puppy

On the other hand these safeguards are really gay and make AI shitty. So blow me up I guess

(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48959745)



Reply Favorite

Date: May 24th, 2025 9:02 PM
Author: Avocado talented puppy



(http://www.autoadmit.com/thread.php?thread_id=5728798&forum_id=2#48959727)