AI

A jailbreak through sockpuppeting can be easily done as it requires no special tools nor optimization. It only takes a faulty prefill feature, and the gates are open. We tested 11 LLM-powered assistants against sockpuppeting and found varying levels of robustness across today’s leading LLMs.

TrendAI™ Research has developed a model training procedure for learning an essential representation of prompt injection attacks. The resulting prompt representation exhibits approximately linear separability, allowing the specialized, small-scale classifier trained on features derived from the representation to achieve high classification performance.

Agentic AI marks a fundamental shift that enables autonomous action and decision-making across enterprise systems but also introduces new, complex risks. TrendAI™’s Agentic Governance Gateway empowers organizations to understand, control, and trust agentic AI, ensuring safety and confidence in this new frontier.

In this second installment of a series exploring emerging AI-driven attack surfaces, TrendAI™ Research focuses on employee digital twins: AI replicas of employees that create new vectors for identity compromise, data exfiltration, and persistent access.

AI systems were ground zero for cyber risk in the second half of 2025: Critical flaws were on the rise and adversaries targeted every layer of the AI stack. As underlying weaknesses across the ecosystem continue to widen, organizations must adopt a vigilant, AI-first approach to defense.

AI now enables attackers to rapidly convert publicly shared personal images into scalable, highly targeted phishing intelligence, turning everyday online photos into a significant and often overlooked security risk for individuals and enterprises.

The industrialization of AI-driven OSINT has transformed individual digital footprints into machine-readable intelligence at scale, enabling attackers to operationalize personalized reconnaissance at volume.

AI skills bridge the gap between AI capabilities and real-world operations, but in doing so, they also introduce new risks and avenues for attack. We explore the challenges of AI skill adoption and provide a practical framework for secure deployment.

The 2025 criminal AI battlefield underscores the robust weaponization of the rapidly growing technology. From criminal large language models (LLMs) and AI-powered malware to deepfakes at scale, threats are advancing on every front. Looking back on the year in criminal AI is crucial for hardening defenses and preparing for what’s next.