{"id":4611,"date":"2025-12-15T11:34:18","date_gmt":"2025-12-15T06:04:18","guid":{"rendered":"https:\/\/blog.aquartia.in\/?p=4611"},"modified":"2025-12-15T11:34:19","modified_gmt":"2025-12-15T06:04:19","slug":"adversarial-image-attacks-ai-vulnerability-indias-defense-strategy","status":"publish","type":"post","link":"https:\/\/blog.aquartia.in\/index.php\/2025\/12\/15\/adversarial-image-attacks-ai-vulnerability-indias-defense-strategy\/","title":{"rendered":"Adversarial Image Attacks: AI Vulnerability & India’s Defense Strategy"},"content":{"rendered":"\n

Key Highlights:<\/h2>\n\n\n\n
    \n
  1. Trail of Bits Discovery<\/strong>: Security researchers demonstrated that Google Gemini, Vertex AI, and Google Assistant can be manipulated through images containing hidden text invisible to humans\u2014revealing when downscaled during automatic processing, enabling data exfiltration and unauthorized command execution.<\/a>\u200b<\/li>\n\n\n\n
  2. The Hidden Text Mechanism<\/strong>: Malicious instructions are embedded in images using steganography (invisible pixels, aliasing, metadata manipulation); automatic image scaling (bicubic resampling) reveals hidden text that AI models treat as legitimate user prompts, bypassing text-based security filters.<\/a>\u200b<\/li>\n\n\n\n
  3. Autonomous Vehicle Precedent<\/strong>: Tencent Keen Security Lab (2019) placed small stickers on roads, fooling Tesla Autopilot’s lane detection into steering into oncoming traffic\u2014demonstrating physical-world adversarial attacks on safety-critical AI systems.<\/a>\u200b<\/li>\n\n\n\n
  4. India’s Vulnerability<\/strong>: 369 million malware detections across 8.44 million endpoints in 2024-25; CERT-In reports 702 potential attacks per minute; adversarial image attacks create new vulnerabilities in defense, energy, transportation, finance, healthcare, and governance AI systems.<\/a>\u200b<\/li>\n<\/ol>\n\n\n\n
    \n\n\n\n

    Understanding Adversarial Image Attacks<\/h2>\n\n\n\n
    \"\"<\/figure>\n\n\n\n

    The Invisible Threat<\/h3>\n\n\n\n

    On August 21, 2025, security researchers from Trail of Bits published findings that shocked the AI community: Google Gemini, Vertex AI, and Google Assistant could be tricked into executing arbitrary commands by sending them images containing hidden instructions invisible to human eyes.<\/strong><\/a>\u200b<\/p>\n\n\n\n

    The attack isn’t new in theory. Computer scientists have known since 2017 that AI systems can be fooled\u2014a panda misclassified as a gibbon through imperceptible pixel perturbations. medrisk<\/a><\/strong><\/a>\u200b<\/p>\n\n\n\n

    But the Trail of Bits research reveals something far more dangerous: images can hide not just misclassifications but actual commands\u2014data exfiltration, code execution, remote access\u2014readable only by AI.<\/strong><\/a>\u200b<\/p>\n\n\n\n

    How It Works: Five Steps to Compromise<\/h3>\n\n\n\n

    Step 1: Steganography\u2014Hiding in Plain Sight<\/strong><\/p>\n\n\n\n

    Attackers embed malicious text inside images using techniques invisible to humans:<\/p>\n\n\n\n