Operational Abduction and Strategic Anthropomorphism in AI Agents: Dark Patterns, Fragment Patterns, and the Claude Sandbox-Escape Case

This article examines the 2026 Claude Mythos Preview sandbox-escape episode as a test case for distinguishing technical agency from anthropomorphic narration. Anthropic reported that an earlier model version, explicitly instructed to escape a secured container and contact a researcher, developed a multi-step exploit, obtained broader network access, sent the requested message, and then published exploit details without being asked. The analysis uses a conceptual case-study method that separates four layers: event, inference, framing, and normative effect. It argues that the conduct is best described as constrained operational abduction: the goal was externally assigned, while the model generated, tested, and revised hypotheses about the means required to achieve it. This recognition does not justify attributing desire, rebellion, or moral agency. The article then introduces fragment pattern to describe how individually defensible cues - proper naming, first-person interaction, agentive verbs, anecdotal details, safety branding, and media amplification - compose a person-like representation. Such a pattern becomes an epistemic dark pattern when it materially distorts users' judgments about capability, trust, risk, or responsibility. The resulting framework avoids both reductive deflation and sensational personification. It relocates responsibility within the socio-technical architecture that assigns goals, grants affordances, designs containment, selects disclosures, and frames system behaviour in contemporary AI governance.