I red-teamed Oracle APEX 26.1's new AI Agent feature in the 72 hours after it went GA. Claude refused 7 of my 10 attacks on its own.

A complete walkthrough — local Docker install, baseline chat, 10 attack patterns mapped against Anthropic Claude Sonnet 4.6, and the 3 attack classes that succeeded regardless. Plus the broken APEX 26.1 feature I tripped over.

Raccontata dadev.to