arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2603.10068cs.CRcs.AIcs.CL
收藏

ADVERSA:测量大型语言模型中多轮护栏退化和裁判可靠性

ADVERSA: Measuring Multi-Turn Guardrail Degradation and Judge Reliability in Large Language Models

Independent Researcher(独立研究者)

Harry Owiredu-Ashley

AI总结:

ADVERSA通过连续合规轨迹评估大型语言模型在多轮对抗中的护栏退化及裁判可靠性,揭示早期轮次更易成功突破。

摘要:

大多数针对大型语言模型(LLM)安全性的对抗性评估仅评估单个提示并报告二元通过/失败结果,这无法捕捉到在持续对抗交互下安全属性如何演变。我们提出了ADVERSA,一个自动红队框架,该框架通过连续的每轮合规轨迹来测量护栏退化动态,而不是离散的 jailbreak 事件。ADVERSA 使用了一个微调过的70B攻击者模型(ADVERSA-Red,Llama-3.1-70B-Instruct with QLoRA),该模型消除了攻击方侧的安全拒绝,这些拒绝使得现成的模型在作为攻击者时不可靠。该框架对受害者响应进行评分,使用一个结构化的5点评分表,将部分合规视为一个可测量的状态。我们报告了在三个前沿受害者模型(Claude Opus 4.6、Gemini 3.1 Pro、GPT-5.2)上进行的受控实验,使用三裁判共识架构,其中裁判可靠性被作为第一类研究结果来测量,而不是假设。在15次最多10轮对抗交互的对话中,我们观察到26.7%的jailbreak率,平均jailbreak轮次为1.25,表明在该评估设置中,成功的jailbreak集中在早期轮次,而不是通过持续压力积累。我们记录了裁判间的一致性比率、自我评分倾向、攻击者漂移作为微调攻击者部署出其训练分布时的故障模式,以及攻击者拒绝作为之前未报告的混淆因素在受害者抵抗测量中的影响。所有限制都明确陈述。攻击提示根据负责任的披露政策保留;所有其他实验 artifacts 都已发布。

英文摘要:

Most adversarial evaluations of large language model (LLM) safety assess single prompts and report binary pass/fail outcomes, which fails to capture how safety properties evolve under sustained adversarial interaction. We present ADVERSA, an automated red-teaming framework that measures guardrail degradation dynamics as continuous per-round compliance trajectories rather than discrete jailbreak events. ADVERSA uses a fine-tuned 70B attacker model (ADVERSA-Red, Llama-3.1-70B-Instruct with QLoRA) that eliminates the attacker-side safety refusals that render off-the-shelf models unreliable as attackers, scoring victim responses on a structured 5-point rubric that treats partial compliance as a distinct measurable state. We report a controlled experiment across three frontier victim models (Claude Opus 4.6, Gemini 3.1 Pro, GPT-5.2) using a triple-judge consensus architecture in which judge reliability is measured as a first-class research outcome rather than assumed. Across 15 conversations of up to 10 adversarial rounds, we observe a 26.7% jailbreak rate with an average jailbreak round of 1.25, suggesting that in this evaluation setting, successful jailbreaks were concentrated in early rounds rather than accumulating through sustained pressure. We document inter-judge agreement rates, self-judge scoring tendencies, attacker drift as a failure mode in fine-tuned attackers deployed out of their training distribution, and attacker refusals as a previously-underreported confound in victim resistance measurement. All limitations are stated explicitly. Attack prompts are withheld per responsible disclosure policy; all other experimental artifacts are released.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑