SecurityHacker Newsabout 1 hour ago
Open-Source Detectors Fail Realistic AI Agent Attacks
5 min read
A benchmark tested 10 open-source prompt-injection detectors against 629 real AgentDojo attacks buried in ordinary tool output. The best caught 51% of attacks at a 2% false-positive rate, while Meta's Prompt Guard 2 caught just 1%. Two detectors flagged 98% of safe tool outputs, showing text-level detection cannot reliably separate attacker instructions from benign data.
Level
Hype check
Tap to vote and see what everyone thinks.
