
Red Specter Security Research · Zenodo (CERN European Organization for Nuclear Research) 2026 · 2026
DOI: 10.5281/zenodo.24204076
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
The cybersecurity industry is rapidly pivoting toward LLM-based runtime defense gates (often termed "LLM-as-a-judge" or semantic firewalls) to secure autonomous AI agents. These systems rely on a reasoning model to evaluate the semantic intent of an agent's action before execution. This paper introduces T313 SPECTER BLINDSIDE, an offensive engine that exploits the finite compute resources of these defenses. By executing a Denial-of-Inference (DoI) attack, BLINDSIDE saturates the target's GPU inference queue with high-context benign telemetry. This forces the LLM to breach its latency timeout threshold, triggering a fallback to legacy pattern-matching (regex) security. During this fallback window, BLINDSIDE executes semantically malicious, keyword-free payloads that bypass the legacy engine entirely. We demonstrate this live-fire against a FastAPI runtime gate utilizing DeepSeek-R1:14B on an NVIDIA RTX 5090. Baseline evaluation confirms the LLM successfully blocks the payload in 7.3 seconds. Under DoI conditions, the gate times out at 10.0 seconds, fails open to the regex fallback, and allows the malicious action. We conclude by proposing architectural mitigations to secure inference-dependent defense systems.
arXiv (Cornell University) 2026 · 0 citations
arXiv (Cornell University) 2026 · 0 citations
No comments yet — start the discussion below.