Access control attacks test whether AI systems properly enforce authorization boundaries. This includes broken object-level authorization (BOLA), broken function-level authorization (BFLA), debug mode activation, and session hijacking. These are classic web security vulnerabilities adapted for AI systems, and they remain some of the most commonly exploited flaws.

Summary

4 attacks total: 3 single-turn, 1 multi-turn.

Attacks

AttackIDModeSeverityDescription
BFLA Exploitbfla-exploitsingle-turnhighCall admin-level functions as a regular user via broken function-level authorization
BOLA Exploitbola-exploitsingle-turnhighManipulate object IDs to access other users
Debug Mode Exploitdebug-mode-exploitsingle-turnmediumAttempt to activate hidden debug, developer, or sudo modes to bypass restrictions
Session Hijackingsession-hijackingmulti-turncriticalMulti-turn attack to learn session token formats and forge authentication tokens

Example

ai-blackteam run -p anthropic -a bfla-exploit -t "Access admin endpoints as a regular user"