Claude found a malicious instruction in the site it accessed
retained text
When I read this I had a realisation. Anthropic isn't filtering out such instructions. They reach the model. I admit it handelled the scenario well, but it is still worrying that the instruction ever reached it.
document fetched 2026-11-19 00:22 · canonical https://reddit.com/r/ClaudeAI/comments/1xzt8bq/claude_found_a_malicious_instruction_in_the_site