FragFuse: Bypassing Access Control of Large Language Model Agents via Memory-Based Query Fragmentation and Fusion
arXiv:2606.15609v1 Announce Type: cross Abstract: Large language model (LLM) agents increasingly rely on long-term memory to support complex task execution, user personalization, and domain adaptation. Meanwhile, emerging access-control mechanisms for LLM agents are being explored to block policy-violating requests and prevent misuse. We reveal a novel attack surface arising from agent memory operations: prohibited content that would trigger access control can be fragmented across interactions, stored in long-term memory in benign-appearing form, and later reconstructed through memory retrieva
The increasing sophistication and integration of LLM agents with memory functions are creating new attack vectors that researchers are actively exploring and exploiting.
This research reveals a fundamental vulnerability in access control mechanisms for LLM agents, potentially allowing malicious actors to bypass security protocols and exploit sensitive data.
Security paradigms for LLM agents must now account for memory-based attacks through query fragmentation and fusion, requiring more robust and context-aware access control systems.
- · Cybersecurity firms specializing in AI
- · Researchers developing memory-safe AI architectures
- · Red-team AI security testers
- · LLM agent developers relying on current access control alone
- · Organizations deploying LLM agents without advanced security audits
- · Users of compromised LLM agent systems
Immediate updates and patches will be required for current LLM agent systems to address this vulnerability.
Future LLM agent designs will need to integrate more resilient memory management and access control from the ground up.
The complexity of securing AI agents could slow their adoption in highly sensitive sectors, pending robust solutions.
This signal links to a primary source. Continuum Brief monitors and indexes it as part of the live intelligence stream — we do not republish source content.
Read at arXiv cs.AI