Lab: Exploiting LLM APIs with excessive agency
Summary: PortSwigger's Web LLM attacks lab demonstrates an LLM-backed live chat whose Debug SQL API accepts a string containing an entire SQL statement, letting the model execute raw SQL on the database. By asking the LLM which APIs it can access and what arguments they take, a user can have it run SELECT * FROM users to reveal username and password columns and a user called carlos, then run DELETE FROM users WHERE username='carlos' to delete that user.
What's compelling: An API that accepts a full SQL statement and is reachable through a chat prompt turns the LLM into a direct path to destructive database commands, and the lab's own note that live LLM responses are unpredictable means the same excessive agency will not behave consistently in production.
Read the original at PortSwigger Research ↗
- Topics
- Vulnerabilities & Patching
- Related
- Attackers Exploit UnrealIRCd Flaw That Enables Server-Level Command Injection Cyber Risks · Nov 14, 2026U.S. CISA adds Check Point and UnrealIRCd flaws to its Known Exploited Vulnerabilities catalog Cyber Risks · Nov 14, 2026PostgreSQL Patches 52 Flaws, Including 22-Year-Old Bug Enabling Authenticated RCE Cyber Risks · Nov 13, 2026