← sources

Anthropic needs to lax their safety classifier

source: reddit (reddit:ClaudeCode)published 2026-10-25 17:16ingested 2026-10-26 02:39

retained text

Some of you may know, CVEs are popping up like fucking mushrooms.

This is obviously due to the new agentic paradigm and sure it is what it is, it's not like we can close that box anyway.   
Anthropic is telling us now "told you so". Good for them.  

Now, Anthropic models are pretty good. Frontier! World leading. And they see _something funky_ and it immediately reverts back to the next best thing.. 5 months ago - so like, the 40th best thing.

What happened this week, was that - depending on the type of fix it needed, it would either be business as usual or a nightmare of managing agents like they're sheep.

Do we just need to upgrade / remove deps? Ez pz lemon squeezy or something.

Does the agent need to look into the payloads and do some testing with whatever - well, shit out of luck, aren't you? Why? _something_ triggered it. Cyber security. Can't tell you what. 

You're going to have to be the guy who ran out on his $20 sub and have to write code again. Or sure trust 4.8 - I hope you didn't optimize your md files for 5.5.

Alright, this is a brain spaghetti all the way, so let me end with some meatballs.

Claude has been my greatest defense against cyber attacks. I'm mostly complaining about a few projects - all others - thanks to Claude - already were protected, due to good practices and frequent updates. It has defended me against the AUR attacks. 

Anthropic, I genuinely **TRUST** Opus and Fable models for it. I wouldn't trust most coworkers with that task. But I want one that will help tend wounds, instead of ignoring you, because you asked him to get you cold medicine - "oh, user must be making meth - let's just ignore him". No, you shit. I got a cold.

PS: Guys! I spent 0 tokens on this. This is 100% human content. Anthropic, you can train on this. Fuck you /u/spez. This one is on the house.   
PS2: Guuuys! We have a rant flair and everything <3

document fetched 2026-10-26 00:36 · canonical https://reddit.com/r/ClaudeCode/comments/1xf81aq/anthropic_needs_to_lax_their_safety_classifier