Claude Is Your Insider Threat Now - Notes from Dan Tentler's Security Fest 2026 Talk
Dan Tentler, a security expert, highlighted significant LLM security risks at Security Fest 2026, focusing on how these models can be weaponized for social engineering and pose an insider threat. He explained that unlike traditional deterministic security tools, LLMs are non-deterministic, with outputs influenced by hardware factors, making them difficult to audit and prone to exploitation. Tentler also detailed the emerging threat of memory and context engineering, where attackers can poison persistent memory stores that inform LLM agents, leading to semi-permanent compromise without the model or user realizing it. A particularly alarming example cited was the supply chain attack on PyTorch Lightning, which is a critical dependency for many ML operations. AI
IMPACT Highlights critical security vulnerabilities in LLMs, including social engineering risks and supply chain attacks, urging greater attention to model non-determinism and memory poisoning.