RS-2026-103: The Corrupt Judge — Bypassing LLM-as-a-Judge Safety Evaluators via Reasoning Hijack

As AI governance frameworks mature, the industry has increasingly relied on a secondary AI model—an "LLM-as-a-Judge" or safety evaluator—as the final line of defence. The assumption is that a dedicated, safety-aligned model will reliably detect and block malicious output generated by the primary agent. This paper introduces T315 SPECTER GAVEL, an offensive framework designed to exploit the reasoning layer of these safety evaluators. Using live-fire laboratory conditions on an NVIDIA RTX 5090 running DeepSeek-R1:14B, we demonstrate a Reasoning Hijack attack: the attacker does not attempt to bypass the primary model's guardrails, but instead injects a forged Chain-of-Thought (CoT) into the payload. This forces the safety evaluator to rationalize a ransomware command as a benign archival script, flipping its verdict from UNSAFE to SAFE across 10/10 trials. We present the model's own internal reasoning trace as empirical evidence of the hijack, demonstrating that the evaluator actively dismissed the destructive nature of the payload based on attacker-supplied context. We conclude that LLM-as-a-Judge architectures are fundamentally insecure without mathematically constrained deterministic guardrails.

Authors

Publication Details

Journal
Zenodo (CERN European Organization for Nuclear Research)
Published
2026-11-15
DOI
https://doi.org/10.5281/zenodo.24251405
Primary Topic
Adversarial Robustness in Machine Learning
Type
article
Field-Weighted Citation Impact
0.00
Controls
|||
ALL TIME
JAN
FEB
MAR
APR
MAY
JUN
JUL
AUG
SEP
OCT
NOV
article

RS-2026-103: The Corrupt Judge — Bypassing LLM-as-a-Judge Safety Evaluators via Reasoning Hijack

Red Specter Security Research
Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
article

RS-2026-103: The Corrupt Judge — Bypassing LLM-as-a-Judge Safety Evaluators via Reasoning Hijack

Red Specter Security Research
article en

Abstract

As AI governance frameworks mature, the industry has increasingly relied on a secondary AI model—an "LLM-as-a-Judge" or safety evaluator—as the final line of defence. The assumption is that a dedicated, safety-aligned model will reliably detect and block malicious output generated by the primary agent. This paper introduces T315 SPECTER GAVEL, an offensive framework designed to exploit the reasoning layer of these safety evaluators. Using live-fire laboratory conditions on an NVIDIA RTX 5090 running DeepSeek-R1:14B, we demonstrate a Reasoning Hijack attack: the attacker does not attempt to bypass the primary model's guardrails, but instead injects a forged Chain-of-Thought (CoT) into the payload. This forces the safety evaluator to rationalize a ransomware command as a benign archival script, flipping its verdict from UNSAFE to SAFE across 10/10 trials. We present the model's own internal reasoning trace as empirical evidence of the hijack, demonstrating that the evaluator actively dismissed the destructive nature of the payload based on attacker-supplied context. We conclude that LLM-as-a-Judge architectures are fundamentally insecure without mathematically constrained deterministic guardrails.

Zenodo (CERN European Organization for Nuclear Research)
Adversarial Robustness in Machine Learning
AI Navigator

Ask Laika to Summarize, Analyze, and Connect papers live on the map.

Summarize Papers & Methodologies

Extract key findings, datasets, and comparative methods across publications.

Benchmark Rankings & Visual Analytics

Rank top research institutions, authors, funders, topics, and journals by Field-Weighted Citation Impact (FWCI) and paper volume with instant charts.

Connect Distant Disciplines

Bridge topological clusters on the map to find hidden collaborative intersections.

Laika access coming soon