safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations.
The message
safeguards flagged this message (https://www.anthropic.com/legal/aup). This sometimes happens with safe, normal conversations.What it means
Anthropic's safety filters flagged your prompt against its Usage Policy, not the cyber or biology rules. The reasoning_extraction detail and the line about safe conversations mark it as a likely false positive, which Anthropic says its filters can produce.
What to do
Double-press Esc to edit your message and retry, or run /model to switch models. If legitimate work keeps getting blocked, report it with /feedback. Turn off automatic switching under Settings > Capabilities if you'd rather it pause than switch.
Claude Code prints this when Anthropic's safety filters stop a request, and the message links to the Usage Policy rather than naming a topic. The model's name sits in front, so on a 1M-context run it reads "Opus 5.5 (1M context)'s safeguards flagged this message". Anthropic's own wording on the next line is the sign it may be a mistake.
- Message links to
- the Usage Policy, anthropic.com/legal/aup
- Detail shown
- [reasoning_extraction]
- Product
- Claude Code
- Anthropic's line
- This sometimes happens with safe, normal conversations
- First fix
- edit and retry, or /model
Why you got it
Anthropic runs detection models over requests and blocks the ones its classifiers read as harmful under the Usage Policy. Its user-safety page is blunt that the classifiers miss:
These features are not failsafe, and we may make mistakes through false positives or false negatives.
The [reasoning_extraction] detail points at the classifier that watches for attempts to pull out the model's private thinking, one of Anthropic's defenses against distillation. Ordinary prompts trip it, which is why the error itself says it happens with safe conversations.
It's worth knowing the check reads more than your last message:
The checks also review everything the model reads, not just your latest message—including memory, content from connectors, web search results, and files, so a block can be triggered by content you didn't type.
How to fix it
Do what the message says first. Double-press Esc to edit your last message and retry, or run /model and send it to a different model. If the block came from something earlier in the session, like a file or a search result, trimming that often clears it.
If legitimate work keeps getting stopped, report it with /feedback. Anthropic says those reports are how it narrows the filters. For defensive cybersecurity specifically, the route is the Cyber Verification Program, which we cover on the related cyber safeguards error page.
Messages that look like it
A version that mentions a cybersecurity topic is the cyber classifier, not this one, and it has its own fix. On Fable 5 and Fable 5.1, a flagged request usually doesn't error at all. It falls back to an Opus model and re-runs, currently Opus 5 for biology and Opus 4.8 for offensive cybersecurity. Opus 5.5 is already near the top of that ladder, so when its own safeguards block a request there's nothing below it to fall back to, and you get this error instead.
You can turn the automatic switching off under Settings > Capabilities, or Config > MODEL & OUTPUT in Claude Code, by toggling "Switch models when a message is flagged". With it off, a blocked request pauses instead of switching.