https://aisafetyresearchgroup.com/risks Documented AI-enabled fraud, deepfake impersonation and agentic exploitation. Nine ranked categories, every figure traced to its primary source. Registry 2026.1. en-us AI Safety Research Group. Figures remain the property of their primary sources. Tue, 27 Oct 2026 12:00:00 +0000 Tue, 27 Oct 2026 12:00:00 +0000 720 https://aisafetyresearchgroup.com/risks/synthesized-voice-family-extortion https://aisafetyresearchgroup.com/risks/synthesized-voice-family-extortion Tue, 27 Oct 2026 12:00:00 +0000 Risk register A cloned parent or child voice stages an emergency and demands immediate payment. A voice can be cloned from under 30 seconds of audio, per the Federal Trade Commission. Record check: no verified case exists of voice synthesis used to carry out a physical abduction. https://aisafetyresearchgroup.com/risks/deepfake-authorised-payment-fraud https://aisafetyresearchgroup.com/risks/deepfake-authorised-payment-fraud Tue, 27 Oct 2026 12:00:00 +0000 Risk register Financial controls defeated by a face and a voice synthesised in real time. Arup authorised HK$200m, about US$25.6m, across 15 transfers after a video conference in which every other participant was synthetic. https://aisafetyresearchgroup.com/risks/ai-automated-spear-phishing https://aisafetyresearchgroup.com/risks/ai-automated-spear-phishing Tue, 27 Oct 2026 12:00:00 +0000 Risk register Reconnaissance and personalisation automated to a cost that makes targeted attacks routine. In a four-arm trial of 101 participants, fully automated campaigns reached a 54% click-through rate against 12% for a control. https://aisafetyresearchgroup.com/risks/synthetic-identity-kyc-bypass https://aisafetyresearchgroup.com/risks/synthetic-identity-kyc-bypass Tue, 27 Oct 2026 12:00:00 +0000 Risk register Liveness checks defeated by injected synthetic video at the point of account opening. US identity fraud losses reached US$27.3bn across 18 million victims, with new-account fraud victims up 31%. https://aisafetyresearchgroup.com/risks/agentic-vulnerability-exploitation https://aisafetyresearchgroup.com/risks/agentic-vulnerability-exploitation Tue, 27 Oct 2026 12:00:00 +0000 Risk register Agents that find and exploit software flaws autonomously. A GPT-4 agent exploited 73.3% of 15 prepared web vulnerabilities with no prior knowledge of them; the same class of agent found and helped close real flaws before use. https://aisafetyresearchgroup.com/risks/non-consensual-synthetic-imagery https://aisafetyresearchgroup.com/risks/non-consensual-synthetic-imagery Tue, 27 Oct 2026 12:00:00 +0000 Risk register Sexualised forgeries generated at scale, overwhelmingly depicting women and girls. A representative three-region survey of 7,231 respondents found 3.2% reporting creation, sharing or threats to share such images. https://aisafetyresearchgroup.com/risks/official-impersonation-voice-spoofing https://aisafetyresearchgroup.com/risks/official-impersonation-voice-spoofing Tue, 27 Oct 2026 12:00:00 +0000 Risk register Authority spoofed at the scale of a phone bank. A US$6m penalty was finalised against the consultant behind a forged presidential robocall, alongside a US$1m settlement from the transmitting carrier. https://aisafetyresearchgroup.com/risks/voice-authentication-spoofing https://aisafetyresearchgroup.com/risks/voice-authentication-spoofing Tue, 27 Oct 2026 12:00:00 +0000 Risk register Voiceprints treated as credentials against attack families that remain unsolved. ASVspoof 5 tests detectors across 32 attack algorithms, and generalisation to unseen attacks is still an open problem. https://aisafetyresearchgroup.com/risks/investment-romance-crypto-fraud https://aisafetyresearchgroup.com/risks/investment-romance-crypto-fraud Tue, 27 Oct 2026 12:00:00 +0000 Risk register The largest reported losses belong to conventional investment, romance and cryptocurrency fraud, where synthesis is an accelerant rather than the mechanism. Reported losses are a floor, not a total. https://aisafetyresearchgroup.com/press https://aisafetyresearchgroup.com/press Tue, 27 Oct 2026 12:00:00 +0000 Newsroom Copy-ready descriptions of the group, headline figures with the agency behind each, citation formats, reuse terms, and what the evidence does not establish.