Realtime AI News
New Survey Maps Full Lifecycle of LLM Vulnerabilities Across Agents, Tools, and Enterprise Deployments
A comprehensive new survey examines security vulnerabilities of large language models across their entire lifecycle, from training to deployment as autonomous agents. The paper highlights how LLMs embedded in enterprise tools, coding environments, and robotic systems face risks that extend well beyond model weights.
A new survey paper on arXiv provides a systematic examination of LLM vulnerabilities across the full model lifecycle and application stack.
The paper notes that large language models are no longer just text generators — they are increasingly deployed in retrieval pipelines, enterprise assistants, coding environments, robotic systems, security operations, and autonomous agents.
These systems can read private data, call tools, write files, execute code, and act across organizational boundaries.
The survey covers attacks, risks, defenses, and open problems, arguing that risks do not arise from model weights alone but from the full deployment lifecycle.
This comprehensive view is timely as enterprises increasingly integrate LLMs into production systems with tool access and agent capabilities.
The research spans security concerns specific to each deployment context — from data leakage in retrieval-augmented generation pipelines to privilege escalation in autonomous agent workflows.
The paper serves as both a taxonomy of known vulnerabilities and a roadmap for security researchers working on LLM safety.
Why it matters
This survey provides a systematic reference for understanding the full landscape of LLM security, offering critical guidance as agentic AI deployments accelerate.