AI agents face prompt injection threats: Risks remain high even after defense upgrades.

[Coin World] A leading AI company recently unveiled an awkward fact: their smart browser product faces an inescapable security issue - prompt injection attacks. In simple terms, it means that hackers can secretly manipulate AI agents to execute malicious operations through instructions hidden in web pages or emails. This is not new, but the company also admits that such attacks are practically impossible to eradicate completely.

To cope with this predicament, they implemented an automated defense system based on large language models. This defense system, trained through reinforcement learning, is capable of continuously simulating new types of attack methods that hackers might use, thereby detecting and blocking vulnerabilities in advance. The results have been positive—recently, it successfully identified and prevented malicious emails that attempted to deceive agents into sending resignation letters.

But there is a key issue here: smart browsers like this essentially possess a certain level of autonomous decision-making capability while also having access to users' sensitive data. With this combination of permissions and autonomy, the magnitude of security risks may not yet reach a level that allows users to feel comfortable using them in their daily lives. In simple terms, the progress in defensive capabilities is not keeping pace with the speed of risk growth.