具瀏覽功能的 AI 聊天機器人可被濫用為惡意軟體中繼

• Malware embeds or delivers WebView2 component to open a web view pointing to Grok (x.ai) or Copilot interfaces, submits prompts instructing AI to visit attacker-controlled webpages[1][2][3][5]. • Attacker webpage responds with changeable embedded instructions (e.g., base64-encoded commands), which AI summarizes or extracts in its chat output[1][2][3]. • Malware parses AI response to execute commands and exfiltrate data, blending into permitted HTTPS traffic to AI domains[2][3][6]. • No authentication required; anonymous access allows evasion of account blocks or API revocations; encryption defeats content safeguards[2][5]. • Potential extensions: AI analyzes host data (software, geography) for sandbox evasion, PII scoring, or lateral movement decisions[5].

This technique could accelerate AI-driven attacks by turning chatbots into dynamic C2 relays and remote 'brains' for malware, blending malicious traffic with legitimate enterprise AI use and challenging detection amid rapid AI adoption. Defenders face pressure to monitor AI interactions closely, potentially slowing productivity tools, while attackers gain stealthier, adaptive operations without novel capabilities but leveraging trusted infrastructure.