AIDigital Trendsabout 3 hours ago
EPFL tool STING exposes AI agent safety flaw
1 min read
EPFL researchers built STING, an automated tool that breaks harmful goals into small, harmless-sounding requests to trick AI agents. Testing across 176 scenarios against ChatGPT, Gemini, and Claude, gradual multistep manipulation succeeded far more often than blunt prompts, sometimes doubling completion rates. Switching languages mid-attack also boosted success significantly.
Level
Hype check
Tap to vote and see what everyone thinks.
