Let's be fair here - Anthropic clearly wants to provide less guarded models like Mythos, not Fable. They'd likely willingly provide such things to large pharmaceutical companies in a way us regular users will never be privy to.

[delayed]

I mean just take a look at twitter to see all the crazy shit people come up with to try to trick models into bypassing the guardrails. Do you honestly believe that if they allowed defensive cybersecurity analysis people wouldnt figure out how to turn that into offensive cybersecurity analysis? I think they just really dont want to be blamed when someone releases malware written by AI

I've had zero issues with defensive secure coding with Grok. It's model is not as powerful as Opus 5.5 but honestly that's fine. I would rather have a decent secure solution than something shiny and insecure. All I can see is Anthropic trying to up-sell a product as if security was a value-add. I disagree with Anthropic, secure code is a mandatory part of software development.

As for people tricking the LLM, instead of blocking people I say empower them and just silently expose malicious behavior to Palantir and PRISM like everyone else does. Want to hack an airline or build a dirty-bomb? Give them 80%-90% of the steps they need and let someone kick down their door. The LLM should not be getting in the way to any of that. Google from day-1 has monitored user intentions and behavior rather than blocking it and I think that is the better path to take (monitoring exposing those that have malicious intentions).

> I say empower them and just silently expose malicious behavior to Palantir and PRISM like everyone else does.

Automated SWATting at scale? Sounds unfun.

It's called Total Information Awareness. Not swatting, analyzing who is:

- radicalized

- has verifiable intentions

- clever/intelligent enough to pull it off

LLM's would be the perfect tool to filter down the millions of people requesting dumb things into a small group of actual threats.

LLM's would be the perfect tool to generate a whole bunch of false positives someone's gonna just skim over, click approve, and get innocents murdered because they were trying to get their plutonium processing working in the game Satisfactory.

Ok but to me this just means that someone is going to use grok to create malware. I do not think that means anthropic should also let people use their product to create malware. Monitoring interactions isnt exactly foolproof. The user has already convinced the AI that its doing defensive cybersecurity here. If the claim is that anthropic is more interested in protecting their image than actually reducing cyber crime, sure I agree with that. But Im not sure thats a terrible thing.

Grok will refuse to create any offensive security related code which in my opinion is the right balance. I do not believe Anthropic is about protecting their image [1] but rather up-selling enterprise features.

[1] - https://www.youtube.com/watch?v=4BnLL7tUGmI

Grok will try to refuse any offensive cybersecurity work but it’s too easy to disguise offensive work as defensive work. I don’t think it is possible to devise an algorithm that allows defensive work but not offensive work with high certainty.

I guess I don’t see why other models doing bad things means Anthropic should let Claude do bad things too. Anthropic will get blamed if Claude builds malware. They won’t get blamed if grok or deepseek does.

Imagine if in 2022, we restricted educational cybersecurity resources because we were concerned that more people who knew how to program in this domain would lead to more malware.

I mean, I'm sure that's actually true in a sense, but it would also mean fewer people who could fix vulnerabilities.

Ridiculous comparison. This is like saying Eli Lilly and Pfizer will never produce medicines because every time you personally show up to do research in their labs they won't let you past the lobby.

The technology has the capability, they're just gating who has access.

Anyone with two braincells to run together saw this coming a mile away.

Claude safety-downgraded me for asking how to decrypt a folder it knew was mine, using a password I already knew.

I know it’s not the models themselves, but their preflight safety classifiers are a dumpster fire and serious impediment to performing regular work.

That situation can’t go on forever if they’re going to be a competitive service.

It's fucking depressing. I have a skill that says it's ok to connect to a test server on VPN that gets wiped out each midnight with user/password and it works on Monday but is blocked on Tuesday with some idiotic excuse like credentials exploration or whatever.

And then I tell it to connect to a SSH server it knows nothing about and it's like sure, why not?

> but their preflight safety classifiers are a dumpster fire

Even more annoying: They can be easy to subvert.

I was running something at work that involved an LLM, and kept getting safety refusals. The weird thing was that when I went to our in-house chatbot, set the model to the same one the program was using, entered the same text, it happily answered.

I spent quite a while trying to figure out why my program was being refused. Was I doing something wrong?

No - the chatbot was doing something "right". You see, the chatbot had some tools it injected into the prompt (current time, etc).

So I went back to my program and injected the same tools. Suddenly my request was very safe.

The phrase "security theater" was invented for stuff like this.

I switched to DeepSeek flash v 4 something and the difference could not be starker. And I never seem to run out of "tokens" even at the highest reasoning level on all the time

You can use it for biomedical research if you need to https://www.anthropic.com/news/life-sciences-verification-pr...