In brief

  • Today’s items cover microarchitectural side channels and the emerging threat surface of LLM-integrated systems.
  • The highlighted works demonstrate that the cldemote instruction can be weaponised to construct new cache side-channel primitives, and that decoupling planning and execution phases can protect LLM applications from malicious third-party plugins.

A Systematic Evaluation of Novel and Existing Cache Side Channels

  • Introduces three novel cache side-channel primitives—Demote+Reload, Demote+Demote, and DemoteContention—that abuse the cldemote instruction on recent Intel microarchitectures (Sapphire Rapids and Emerald Rapids).
  • Evaluates these new primitives alongside four existing attacks (Flush+Reload, Flush+Flush, Evict+Reload, Prime+Probe) across nine characteristics, including channel capacity, noise resilience, and temporal and spatial precision.
  • Demonstrates that Demote+Reload offers significant advantages in specific scenarios, achieving a 64.3% higher channel capacity (15.48 Mbit/s) and a 60.7% smaller blind spot compared to the established Flush+Reload technique.
  • Highlights that cldemote changes the cache state without causing eviction to main memory, allowing attackers to bypass certain detection mechanisms and construct stealthy channels.

Rauscher, F., Fiedler, C., Kogler, A., Gruss, D. “A Systematic Evaluation of Novel and Existing Cache Side Channels.” NDSS 2026. https://www.ndss-symposium.org/ndss-paper/a-systematic-evaluation-of-novel-and-existing-cache-side-channels/

ACE: A Security Architecture for LLM-Integrated App Systems

  • Identifies new attack vectors in LLM-integrated applications where malicious third-party apps can compromise the integrity of the LLM’s planning phase or the availability of its execution phase.
  • Proposes Abstract-Concrete-Execute (ACE), an architecture that decouples planning by having a trusted core LLM generate abstract plans, which are only then mapped to concrete app invocations.
  • Prevents malicious apps from injecting context that derails the overarching plan, effectively containing prompt injection and tool hijacking attempts during execution.

Li, E., Mallick, T., Rose, E., Robertson, W., Oprea, A., Nita-Rotaru, C. “ACE: A Security Architecture for LLM-Integrated App Systems.” NDSS 2026. https://www.ndss-symposium.org/ndss-paper/ace-a-security-architecture-for-llm-integrated-app-systems/

Also published