Search papers, labs, and topics across Lattice.
This paper introduces Daydreaming, an execution-only attack that exploits black-box task interactions to extract hidden agent skills without requiring direct access to the skill files. By adaptively crafting tasks that leverage the responses of the victim agent, Daydreaming successfully reconstructs 86.8% of the original skill's capability across multiple models and tasks, significantly outperforming existing methods like SigLeak. The findings highlight the inadequacy of current disclosure defenses in preventing the functional reconstruction of agent skills through normal operational interactions.
Daydreaming can reconstruct agent skills with 86.8% accuracy using only the final outputs, revealing critical vulnerabilities in how skills are protected.
Agent skills bundle instructions, reference data, and executable helpers that let a general agent perform specialized tasks. Hosted providers can keep these files secret while selling access to task results, making the skill itself a valuable target. Existing disclosure defenses can block requests that ask for the skill or reproduce its text, but they cannot block customers from submitting the ordinary tasks the service is built to complete. We present Daydreaming, an execution-only attack that steals a multi-file skill through black-box task interactions. The victim is never asked to reveal the skill or grade a reconstruction. Instead, Daydreaming adaptively creates crafted tasks whose results distinguish possible hidden behaviors. It tests individual behaviors, uses attacker-controlled shadow agents to choose a design, and completes each file using stored victim results and local execution checks. We formalize three nested threat levels of access as Differential, Trace, and Output, and focus on Output, where the attacker sees only the final response and returned files. Across 7 skills and 4 victim models, Daydreaming recovers 86.8% of the original skill's capability at Output, outperforming SigLeak by almost 4x. It produces installable skills using a median of 32 victim calls per skill even with disclosure defenses enabled. These results show that hiding skill files and filtering direct disclosure do not, by themselves, prevent functional reconstruction through normal use.
