Paper 2025/1774 Original: eprint.iacr.org/2025/1774 ↗

Adaptive-Controlled Mutual TLS for Large Language Model Systems

Abstract

Mutual Transport Layer Security (mTLS) under- pins authenticated, confidential communication across modern service meshes, but its deployment stance in machine-learning platforms is typically static—fixed cipher suites, certificate life- times, and re-authentication schedules chosen for worst-case threats rather than observed risk. Large Language Model (LLM) serving pipelines exacerbate this rigidity: traffic is bursty, topolo- gies reconfigure dynamically under autoscaling, and sensitive artifacts such as prompts, training features, and evaluation data traverse heterogeneous substrates. In this paper we argue that mTLS for LLM systems[1] should be governed by adaptive control rather than static policy. We formalize a feedback loop that ingests multi-modal telemetry—connection error codes, handshake latencies, anomaly scores from request semantics, workload attestation freshness, and service-level objective (SLO) drift—and outputs fine-grained adjustments to transport posture: client-certificate renewal cadence, certificate path length and key type selection, session resumption eligibility, early data gat- ing, proof-of-possession challenges, and revocation propagation thresholds. The controller targets two coupled objectives: maintain cryptographic assurances (mutual authentication, forward secrecy, and replay resistance) while bounding the cost of security on tail latency and throughput during high-load inference.