newsfilter
← Interviews
AI Engineer7 min

Your KV Cache Can Spot Attacks on Your AI Agents - Ari Azarafrooz, Intrinsec AI

Ari Azarafrooz

Conference Presentation, Keynote
  • AI agent application failure modes are projected to expand due to prompt injections, jailbreaks, broken tool calls, data exposures, hidden backdoors, and drift.
  • Security strategies propose integrating Key-Value (KV) cache data with existing guardrails to detect signals that bypass traditional filters because they still become computation.
  • A three-step implementation plan involves establishing a baseline for AI workloads, identifying rare knowledge base (KB) pages against that baseline, and classifying or explaining resulting anomalies.
  • Anomaly detection and classifiers are planned for direct integration into inference engines using libraries such as LMCache to harden policies based on computational anomalies.
  • Real-time alerts derived from computational anomalies are expected to accompany future tool calls and policies, with the capability to identify specific page sections where anomalous events occur.
  • Anomaly detection systems anticipate flagging approximately 70% to 80% of alerts as meaningful and useful, capable of identifying anomalies in voice agents where no text exists for filtering.
  • Trained models utilizing KV pages as features are predicted to correctly generate information regarding personal details, home addresses, locations, and travel data despite not being trained on audio.
  • Production environments face challenges regarding model changes, fine-tuning, and new releases, creating difficulties in maintaining shared baselines as different agents write distinct coordinates across varied models and routers.
  • Learning separate baseline classifiers and explainers for every model is deemed computationally prohibitive, necessitating a mapping of all model-specific data into a single shared representation space.
  • Real-time analysis currently faces delays due to the slowness of moving large data objects, such as 1,256-token chunks approximating 40 megabytes of KV compression, though specific solutions claim to resolve these latency and data movement issues.
  • Teams are encouraged to treat security as a foundational design decision when constructing their own inference stacks to address these evolving security and performance demands.