What happened
Anheng Information CTO Liu Bo said the speed of AI iteration has become the industry's central security challenge, noting that most AI agents today are still used only in personal or small-scale experimental settings. Applications that truly run inside production systems, he added, are just beginning.
Liu Bo outlined two emerging areas of security risk as data moves through model training, inference calls, and agent execution. One is the technical difficulty of defending against external attacks in this new environment; the other involves privacy issues around data infrastructure and high-quality datasets, especially under requirements such as data being 'usable but invisible' and 'usable but unavailable.'
In response, Anheng is building an 'AI plus dynamic data security' approach instead of relying on static rules. By combining AI with UEBA, the company says it can automatically label and identify data, locate sensitive or important data, and determine who is using it and whether those actions are compliant.
Why it matters
Liu Bo's central concern is that AI technology is evolving much faster than security systems can follow. Agent frameworks are still changing, and risks range from malicious operations and data poisoning to simple operational mistakes. Long chain-of-thought tasks can cause agents to fail or act incorrectly, while attackers may also take control of agents for malicious purposes.
Because the technical framework for agents has not yet stabilized, security vendors must respond quickly and deliver new solutions for each new risk. Liu Bo stressed the need for rapid iteration before AI agents become widely deployed.
Key facts
Anheng Information CTO Liu Bo identifies rapid AI iteration as the core security challenge.
Most AI agents are still in personal or small-scale experimental use, with production deployments just beginning.
Anheng proposes an 'AI + dynamic data security' system, using AI and UEBA to automate data identification, labeling, and compliance monitoring.
Liu Bo predicts broader AI agent adoption will occur between 2027 and 2029.
What to watch next
Liu Bo expects AI agents to enter larger-scale applications around 2027 to 2029. Until then, he says security providers should offer an overall AI safety plan that addresses major problems before customers launch agents, then keep updating it as agent technology and risks evolve.