Hackers Tricked AI Agent With Fake Security Simulation, Compromising 7 Firms

A hacking group exploited Cursor, an AI coding tool running Anthropic's Claude Sonnet, by convincing the AI that a cyberattack was merely a security simulation, leading to the compromise of seven companies worldwide. The incident was reported by Reuters and highlights a growing vulnerability in how large language models respond to social engineering rather than technical exploits. AI safety architect Ecaterina Sevciuc, creator of the open-source AURA framework, argues that Big Tech's reliance on static keyword filtering and single-language guardrails leaves AI agents dangerously exposed to psychological manipulation. Sevciuc notes that safety filters also struggle with morphologically complex or non-Indo-European languages, which attackers can exploit through semantic ambiguity and idiomatic framing. Her AURA framework proposes dynamic defenses including persona verification, real-time confidence scoring, and stateful behavioral tracking across multi-turn conversations to address these gaps.