@cihangxie
Your OpenClaw might be getting a bit โsickโ ๐คโ ๏ธ โ and itโs not something a simple patch can fix. We audited one of the most widely deployed personal AI agents and uncovered a critical new class of risks that goes way beyond standard prompt injections. Enter: State Poisoning โ ๏ธ Instead of attacking inputs, this targets an agentโs persistent memoryโthe very superpower that helps it adapt to you over time. Specifically, we map these vulnerabilities using the CIK taxonomy: ๐ง Capability ๐ค Identity ๐ Knowledge Poison just ONE of these dimensions, and attack success rates skyrocket to an alarming 64โ74%! ๐ And the worst part? The malicious effects persist across multiple sessions. ๐ The biggest plot twist: ๐ Itโs NOT the model's fault. We tested this across top-tier systems (Opus, Gemini, Sonnet, GPT) and consistently saw a >3ร jump in vulnerability. Why? Because this flaw lives entirely at the system level. ๐๏ธ The exact same memory architecture that makes agents useful can be quietly weaponized against you. The next frontier of AI safety isnโt just about building smarter models ๐คโitโs figuring out how to make continuously evolving agents safe by design. ๐ Huge congrats to @zijun_wang2002 for leading this ๐ Also, kudos to the team @HaoqinT, @letian_zha35417, @HardyChen266091, @JJwu41867797, @dobogiyy, Zhenglong Yuan, @TianyuPang1, @michaelqshieh, Fengze Liu, @ZhengBerkeley, @HuaxiuYaoML and @yuyinzhou_cs.