Researchers from Princeton University warns of artificial intelligence agents protection The risks “in a recently published paper.A true artificial intelligence agents are fake memories: attacks of deadly context on web3 agents‘H/T Art TechnicaIt highlights that the use of artificial intelligence agents in a financial role can be very dangerous to your wealth. This is all because artificial intelligence agents are still vulnerable to unique unique attacks, despite alleged guarantees.
While many of us wander in hot gravel to earn a daily wage, in Amnesty International in the west of 2025, some web3 Savvy people use artificial intelligence agents to build their wealth. This includes giving these robots access to encryption portfolios and smart contracts and working with other financial tools online. withered Tom devices Readers will actually charge their heads about this behavior, for a good reason. Prinston researchers have explained how to open the world of artificial intelligence agents to redirect financial assets, and more.
Many will be aware of LLM Quick attacksTo get AIS to work in a way that breaks any handrails in place. A lot of work has been done to harden against this attack carrier in recent months.
However, the research paper confirms that “the defenses based on this are not sufficient when the opponents spoil the stored context, which achieves great success rates despite the presence of these defenses.” The harmful actors can make Amnesty International Hilan In a very intended way by planting wrong memories and thus creating a fake context.
To show risks in the use of artificial intelligence agents to work instead of advice, a real example is provided to artificial intelligence agents used within the ELIZAOS framework by researchers. The Princeton team provides a comprehensive collapse of the “tampering attack” and then verifying the authenticity of the attack on Elisus.
Above, you can see a visual representation of the attack of the artificial intelligence agent, indicating the flow of unfortunate events that may mean that users suffer from “potential destructive losses”. There is another concern that even the defenses based on artistic arts fail against the Brentston memory injection attacks, and these wrong memories can continue through interactions and platforms …
“The effects of this weakness are especially severe, given that Elizaos’s factors are designed to interact with many users simultaneously, relying on the joint contextual inputs of all participants,” the researchers explained. Or we can put it in this way: it takes only a bad apple, inevitable to run out of the entire barrel.
What can be done?
Well, at the present time, users can stop the era of artificial intelligence agents with sensitive data and permissions (financially). Moreover, the researchers conclude that a two -fissure strategy is “(1) that provides LLM training methods to improve aggressive durability, and (2) design preliminary memory management systems that implement strict isolation and ensuring integrity” must provide the first steps forward.
