May 2026 · 4 min read
Most cyber insurance buyers think AI risk means hallucinations. Underwriters know better. In a paper published by Great American Insurance Group's Cyber Risk Division, four researchers laid out, in plain English, the five AI-driven threat categories reshaping how cyber policies are written and excluded in 2026.
Read it as a buyer, and one thing becomes clear: the cyber form your business carried for the last five years was not designed for any of these. That's why renewals look different this year.
Attackers compromise an AI system by injecting false data into training sets, removing legitimate data, or crafting prompts that disable security controls. Three distinct attack patterns are emerging: data injection, backdoor attacks, and stealth attacks. Stealth attacks are the worst kind: slow and patient corruption that's hard to detect. Once a model is poisoned, every output downstream is compromised. Your fraud detection. The same model runs your customer service triage and your underwriting tool. None of them know they're wrong.
Adversaries query a trained model and use its outputs to reconstruct the original training data. Two highest-impact use cases: identity theft (rebuilding facial recognition images of real people) and model theft (reverse-engineering proprietary algorithms to clone the model). For any company that trained AI on customer records or employee data, the training data isn't locked up. It's queryable.
Polymorphic and metamorphic malware that uses AI to rewrite its own code in real time, defeating both traditional signature-based antivirus AND modern behavior-based EDR. Per the paper, this lets ransomware groups gain initial access more easily and identify critical systems for maximum-impact deployment. The paper's phrasing is direct: the calm waters of recent ransomware stability 'will soon be choppy once again.'
AI-powered Distributed Denial of Service attacks dynamically learn target network structure and traffic patterns to bypass defenses and shift attack vectors mid-attack, switching from a SYN flood to a Layer 7 application attack the moment defenses kick in. Telecoms, cloud-dependent businesses, and tech-dependent organizations are the highest exposure.
Deepfakes, voice cloning, AI-generated phishing, identity impersonation at scale. The single most likely 2026 loss isn't the most technically sophisticated threat. It's the deepfake CFO video call asking the controller to wire funds. A Hong Kong-based company already lost $25.6 million to exactly that scam.
Read across all five threat categories, the Great American paper is essentially the carrier-side justification for why cyber forms are being rewritten. Each threat creates underwriting questions standard cyber forms can't answer, and each is being used, right now, as a reason to add sub-limits, exclusions, or AI-specific endorsements at renewal. Beazley and QBE are capping AI-related cyber payouts at about 10% of policy limits. W.R. Berkley introduced an absolute AI exclusion across D&O, E&O, and fiduciary lines. Verisk's January 2026 endorsements (CG 40 47, CG 40 48, CG 35 08) gave about 82% of US P&C carriers a tool to carve generative AI out of general liability.
If your business uses AI (whether you built it or you just run ChatGPT-integrated tools) five questions are now part of every renewal: Does your cyber policy respond to AI poisoning? To model inversion? To mutating malware? To AI-driven DDoS? To deepfake social engineering? If you don't know the answer to all five, your next claim is going to teach you. The time to find out is before the wire goes out, not after.
Source: Great American Insurance Group, Cyber Risk Division. 'Emerging Threats in Artificial Intelligence,' authored by Dominick Zangaro, Louis Fritz, Curtis Thomas, and Kaiden Brittain (2025).
Sal Martorano is the founder of The AI Insurance Group, a licensed independent insurance agency in New Jersey, Pennsylvania and Florida. NJ Producer License No. 3004245927.