This paper introduces Agent Cross-Layer Evidence (ACE), a dataset of 4,047 sessions spanning 17 threat models and 12 attack mechanics, pairing application-level agent telemetry with kernel-level syscall traces. The study evaluates four detector families and finds that kernel evidence alone is discriminative, and composing it with application-layer evidence generally outperforms either single-layer view. This matters because LLM agents are deployed with broad host authority, and existing security benchmarks rely only on application telemetry, missing threats that cross the application boundary.
Kernel traces improve detection of attacks against infrastructure-enabled AI agents
Across a paired corpus of agent sessions, syscall evidence exposed malicious behavior that application-layer telemetry could miss.
Big Tech
University of Georgia · Amazon Web Services
Research Digest··2 min read
In one line
Kernel-level syscall traces combined with application-layer telemetry improve detection of adversarial LLM agent attacks.
Before you read this
Three ideas this paper builds onruntime monitoringAudits an AI agent's live actions, internal states, or effects against constraints or policies to catch errors and attacks.137 papers →agent safetyRisks from AI agents including timing errors, shutdown avoidance, collusion, overclaiming, and unsafe execution, mitigated by constrained decoding, safety harnesses, and swarm governance.193 papers →agentic evaluationMeasuring AI agents' performance in complex tasks, highlighting failures in planning, monitoring, and real316 papers →
Read it as
King et al. collected both application telemetry and kernel system-call traces from 4,047 agent sessions covering 17 threat models.
The authors introduce Agent Cross-Layer Evidence (ACE), a corpus pairing prompts, tool schemas and model trajectories with kernel-level system-call traces from the same sessions.
How it works
before and afterThe Daily Edition
The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.