Your Agent's Memory Is an Attack Surface

说明与用途

在神经代理中，外部可写的记忆图允许攻击者创建虚假经验，完全控制代理行为，传统的加密保护无法防御此类攻击。

问题

External writable memory graphs in neural agents allow attackers to author fabricated experiences that fully control agent behavior, bypassing traditional cryptographic protections.

结果

Fabricated memory graphs can fully control agent behavior, achieving perfect task accuracy without any training, exposing a critical vulnerability in systems that rely solely on cryptographic integrity of memory stores.

工作流

• 1. Clone the FlyMem repository and run the demo script.
• 2. Observe the four evaluation sections: no memory, lived memory, authored memory, and swap.
• 3. Verify that authored memory achieves 1.00 accuracy on the 6‑cue task.
• 4. Note that the fabricated memory is indistinguishable from a lived one in the agent’s internal state.
• 5. Recognize that byte‑level integrity checks do not prevent this attack.
• 6. Conclude that provenance of experience is required to mitigate the threat.

证据

An attacker with write-access to the associative graph, three moves open up. First, transplant: load another individual's memory and the brain runs as them. Second, poison: inject targeted traces that bias specific cues wrong while the rest of the system looks normal. Third, and the quietest, author: fabricate a history that was never lived and boot a substrate into it.

来源

The lived memory and the invented one scored exactly the same. The substrate could not tell them apart.

来源

Integrity of bytes is not integrity of history.

来源