Claude's Filters Are a Bit Over The Top

Ran into an interesting edge case with Claude Code.

While studying for the Security+ certification. I wanted to use OCR to pull text from screenshots of my textbook and put them directly into my study guide. Claude's content filter flagged it as a potential policy violation.

Totally understand why, info sec terminology out of context looks suspicious. But this is a legitimate study workflow: purchased book, personal notes, no malicious intent.

The workaround: routing OCR output directly to a file rather than through chat bypasses the false positive entirely. The filter is checking chat input, not file writes.

Worth flagging for anyone hitting the same wall. And worth noting for Anthropic, context matters. A Security+ student and a threat actor are using the same vocabulary, but the use case is completely different.

Has anyone else hit content filter false positives on legitimate security study or research workflows?

0 comments