"Blind Compliance": Coercive Control in AI Jailbreaking

arXiv:2611.08226v1 Announce Type: new Abstract: The explosion of generative AI systems and concerns surrounding their rapidly scaling capabilities has led to an adversarial cycle between increasingly narrow content restrictions ("guardrails") and increasingly sophisticated structured prompting techniques ("jailbreaks"). Technological frameworks surrounding such jailbreaking remain largely focused on the system's output: whether harmful or disallowed content is ultimately produced for the user. However, this approach, which draws largely from classical cybersecurity, does not address the uniquely relational dimension introduced by generative AI: in many cases the user is not just accessing a technology via an exploit, but is in conversation with it as part of the exploit. Here we present an alternative lens which centers the interaction rather than the output. Drawing on interpersonal violence scholarship, we present a novel framework of coercive control categories with a particular focus on aspects which are embodied versus disembodied. We then map how these disembodied coercive control dynamics are structurally embedded in modern AI training paradigms, and how those embeddings are exploited via a particular suite of persona-based jailbreaks. In this lens, we reframe AI safety and alignment as not merely a technical security problem, but as another surface upon which we see enacted the relational dynamics of power and control.

Introduces coercive control dynamics in AI interactions, implying potential harm to users. Reframes AI safety as a power and control issue.

The paper highlights that AI jailbreaking reflects coercive control dynamics embedded in AI systems, indicating risks of relational power imbalances and manipulative interactions beyond technical concerns. This suggests material societal harm related to manipulation and control in human-AI interactions.

This is a conceptual academic framework paper proposing a new lens on AI jailbreaking vulnerabilities rather than reporting a measured incident or deployed harm. It implies existing safety paradigms have a structural blind spot, which is mildly concerning but speculative and theoretical rather than evidenced by concrete exploitation data.

This paper proposes a new framework to analyze AI jailbreaking through the lens of coercive control, drawing parallels with interpersonal violence. It claims that AI training embeds 'disembodied coercive control dynamics' which are exploited by jailbreaks. The implication is that current AI safety focuses are too narrow and miss a crucial, deeply negative, relational dimension of AI's societal impact.

The text describes a cycle of control and evasion in AI systems. It frames AI interactions as coercive, with users exploiting relational dynamics. This implies significant risks to user autonomy and safety.

SOURCE     arXiv cs.CY (Computers and Society) (tier 1)
   ↓
DOCUMENT   dbcd4b8f-dd62-4bfa-8434-58862f3d700e
           https://arxiv.org/abs/2611.08226
   ↓
EVIDENCE   2 extracted excerpts
   ↓
MODEL RUN  5 runs, methodology 1.0.1
   ↓
SCORE      -56  (Adverse)
   ↓
CONFIDENCE 60%