Claude gets halted with “Bad request. Output blocked by content filtering policy…” increasingly frequently.
There is no rhyme or reason why this might be happening - I’m not doing anything sketchy - this is regular development work.
It’s happing across several completely different and totally unrelated projects.
Steps to Reproduce
Unknown - there must be some kind of aggressive and inappropriate policy rule that’s been introduced in the last few days that needs to be rolled back, or, we need to be given some way to bypass this show-stopping block.
Expected Behavior
We should be warned, if at all, about this - NEVER blocked, and ABSOULTELY NEVER should the agent be had-stopped when this happens!
We should also be provided with a detail reason - just saying we’re blocked, time-and-again, with no reason is infuriating!
This error is related to Claude’s content safety filters. This can’t be turned off on Cursor’s side because the filtering happens at the model provider level.
Could you share a bit more context so we can look into it?
What kind of code or project were you working on when this happened?
Can you reproduce it with a specific prompt?
For now, you can try switching to another model, like Claude Sonnet or GPT‑5, to keep going.
Every time this happens, I notice it’s because the agent is trying to output something that would make more sense to be copied from somewhere, instead of generated on-the-fly.
You need to REMOVE whatever filters you’ve got for this kind of thing, and not mix that up with safety-related stuff.
It’s not your concern how we generate things like our license text files !!
@deanrie - this is a serious mistake - please can you escalate this to your engineering team to fix - if it’s hitting me repeatedly, and always a mistake, it’s for-sure ruining the day of many other users too!
Since this is happening at a Claude level and not Cursor and no setting in Cursor can turn this off, it isn’t a bug, you just need to tackle this from a different angle. Safety measures are in place for reasons I don’t need to explain.
You need to REMOVE whatever filters you’ve got for this kind of thing, and not mix that up with safety-related stuff. - You need to understand Cursor is a shell giving you access to the large array of AI, there is no way to turn off a filter.
You could however tackle this differently so you achieve what you’re trying to do. Think outside the box.
good idea - I could program my “stop” hook to check for those, and attempt a “you got blocked; re-word before re-try” autoprompt! Thanks for nudging my brain that way
I’m not sure this is related to the model - it looks like a network-level block that’s checking the output from the model - hard to say if this is between anthropic and cursor, or between cursor and us: but I’d guess the latter, because it makes zero sense for Anthropic to be blocking LICENSE.txt wording at the network, but Cursor have, ahem, somewhat of a track record for gung-ho updates with no testing… so it’s highly conceivable they’re the culprit here…
Can confirm that this error comes directly from Anthropic. You can read about it here:
Not aware of any other user reports of hitting this unexpectedly. But @Chris_Drake if you fall into a situation where it easily reproduces and you can provide a request ID with privacy mode disabled (Cursor Settings > General> Privacy) happy to do a sanity check that there’s nothing in our harness that, combined with your prompt, is triggering this.
