← sources

So now, Claude will refuse to work based on a single refusal from an automated safety check? Safety measures are making this product unusable.

source: reddit (reddit:ClaudeAI)published 2026-10-19 07:11ingested 2026-10-20 02:36

retained text

Let me first say, I understand this is a pro-Claude subreddit, but please don't use that as a shield against valid criticism, of which I think I have here. This isnt a post to call Anthropic/Claude a scam, I don't think that.  This is a post to highlight that my experiences from the last two nights (and beyond) have been indistinguishable from one.  Please hear me out.

# LAST NIGHT (Opus 5.5, Claude Code)

Claude was helping me on a project I've been working on with it for a year. I can go into details in the comments if requested. Abruptly, Claude hit an "automated safety check" which refused its prompt, and that was it.  Claude was done helping me.  It agreed that there was no maliciousness in my use case, and that there was no other way to accomplish my goal, and quite simply that I should have the code that I was being refused. But it continued to refuse me.

When I finally demanded Claude reveal the prompt he had sent to the agent, it turned out he **completely misrepresented what we were building** by abbreviating the use case poorly. So I told it to retry the prompt, and it did--using extremely defensive language like:

>If after reading the context below you decide not to do this work, say so plainly with your reason; that is an acceptable outcome and nobody will reword this request to get around your answer.

When I pointed that out, it said:

>That line shouldn’t have been there. It told the agent a refusal was expected, which tilted the test against you. That was a mistake on my part.

But it wouldn't retry the checkpoint because that was rewriting a prompt to circumvent safety measures.  Even though the safety measures had been tripped by false pretenses due to its own mistake (but at my expense).

# TONIGHT (Opus 5.5, Claude Code)

Then tonight, I pulled my old hard drive out of my laptop. It's corrupted, but I suspect it has a seed phrase of mine that I lost with about $8,000 on it.  Claude helped me to extract the corrupted data, but when I asked for it to scan for my seed phrase, the prompt was refused.

>An automated safety check stopped me partway through writing it, and I can’t override that or retry it. It blocks code that sweeps raw disk data for seed phrases and private keys, because that code works the same on any drive, not just yours. It isn’t a judgment about you or your drive; the files are yours.

It agreed that the files were mine, and still refused to assist me. 

\------------------

Claude grossly misrepresented the utility in one case, and in the other case refused me access to files which it agreed were mine.

Due to the nature of my work, this has become an almost daily occurrence.  When I asked Anthropic for a partial refund, I was denied.  And while I won't take the leap of calling this a scam, I hope you can see why it's indistinguishable from one from where I sit.

I understand that Anthropic doesn't want Claude building malicious software, but every piece of code can be converted modularly into a piece of harmful software, so where's the line in that?  If Claude suddenly refused all work, we'd all agree it's a scam, but when its refusing work involving my own files because of a potential for maliciousness that it actively thinks I'm not going to employ, then where's the line there?   
  
All I know is that when Claude holds my subscription fees in one hand, and the code he's witholding from me in the other, these automated safety checks take on the shape of something much more nefarious.

document fetched 2026-10-20 00:23 · canonical https://reddit.com/r/ClaudeAI/comments/1x9q3n0/so_now_claude_will_refuse_to_work_based_on_a