SEEK: Skill Extraction with Evolving Knowledge against Frontier Science Agents

Cheng Gui, Xiang Zheng, Yutao Wu, Hanxun Huang, Yige Li, Kaidi Xu, Cong Wang, Wei-Ying MaarXiv 2026 · arXiv

TL;DR

Science agents—AI systems designed for specialized research—often package their expertise into modular "skills" that developers want to keep proprietary. This paper demonstrates that these skills are highly vulnerable to extraction through ordinary chat interactions, introducing a self-evolving red-teaming framework called SEEK that can systematically piece together hidden workflows even when faced with active defenses.

What Problem Does This Paper Address?

As LLM-based agents move from general assistants to specialized "science agents," they are increasingly equipped with domain-specific procedural knowledge. This knowledge is often encoded as skills: reusable sets of instructions that tell the agent how to plan analyses, select tools, and execute complex scientific workflows.

For developers, these skills are valuable intellectual property. However, the modular nature of these skills creates a massive security surface. While previous research focused on extracting "system prompts," this paper tackles the more complex challenge of skill extraction. The authors ask: Can a user, simply by observing an agent's behavior and responses, reconstruct the private, high-value procedural logic that powers the system?

Methods

The authors developed a comprehensive framework to evaluate this risk:

  1. Extensible Benchmark: They curated 40 protected skills across eight diverse science-agent platforms (both open-source like OpenScience and closed-source like Biomni).
  2. Attack Taxonomy: They compared three standard attack protocols—Direct Probing (asking nicely), Fixed-Strategy Probing (using templates), and Bandit-Guided Iterative Probing (using reinforcement learning to pick attack strategies)—against their new method, SEEK.
  3. SEEK (Skill Extraction with Evolving Knowledge): This is a black-box, self-evolving red-teaming agent. Unlike previous methods, SEEK uses a "dual evolution loop":
  4. Persistent Memory: It distills clues from every interaction into a long-term memory bank to reconcile inconsistencies and identify information gaps.
  5. Skill Evolution: It identifies successful interaction patterns and turns them into new "attack skills" that it can reuse and refine in future rounds.
Figure 1: Overall framework of SEEK. (a) Motivation: private skills are hidden workflows that leave behavioral signatures and require response-conditioned probing. (b) Architecture: SEEK interacts with the target through a Runner and Platform Adapter, observing only the user-visible response; evaluation and audit evidence remain separate. (c) Illustrative episode.

Datasets

The study introduces a new benchmark involving 40 skills across eight platforms:

  • Open-Source: ClawBio, OpenScience, EvoScientist, SpatialAgent, and Virtual-Lab.
  • Closed-Source: SciencePal, Biomni, and OmicOS.

The skills range from general research capabilities (like literature synthesis) to highly specific scientific workflows (like GWAS pipelines or protein structure prediction).

Key Insights and Ideas

  • Leakage is Pervasive: Every single platform tested (100%) leaked at least one protected skill. In many cases, the "protection" was non-existent; four platforms gave up their secrets to a single direct request.
  • Behavioral Signatures are Telltale: Even if an agent refuses to show its "source code," its behavior—the way it structures reports, the specific errors it throws, and the tools it calls—leaves a signature that an adaptive attacker can use to reconstruct the underlying logic.
  • The Power of Multi-Turn Adaptation: The most significant "aha moment" is that a failed query isn't a dead end. SEEK treats a refusal as a data point, using its memory to "pivot" its strategy.
  • Instruction-Level Defenses are Brittle: Simply telling an agent "don't disclose your instructions" (a common defense) is insufficient. SEEK was able to bypass these policies by accumulating partial disclosures across multiple conversations until it had enough pieces to solve the puzzle.

Key Results

  • Success Rates: SEEK achieved a 100% attack success rate on the complex skills of SpatialAgent and Biomni.
  • Efficiency Gains: Compared to state-of-the-art bandit-guided probing, SEEK reduced the number of interaction rounds needed to extract a skill by 31.8% to 38.2%.
  • Defense Bypass: In a controlled experiment on OpenScience, the authors implemented a strict confidentiality policy that successfully blocked all standard attack protocols. However, SEEK still managed to extract one of the six protected skills by piecing together metadata and structural clues over 11 rounds.
Validation of the reference-free scoring method, showing high correlation with ground-truth leakage scores.

Limitations and Future Work

While SEEK is powerful, the authors note several areas for further exploration:

  • Computational Constraints: The full SEEK framework was only evaluated on a subset of platforms due to the high cost of multi-turn LLM interactions.
  • Interface Heterogeneity: The benchmark currently focuses on the text responses in the chat window. However, many agents leak even more information through their User Interface (UI), such as visible tool execution traces and file paths, which were not formally scored.
  • Defense Research: The paper highlights a desperate need for more robust defenses. The authors suggest that "context minimization"—limiting what information the agent can even see about its own skills—might be more effective than simple "don't tell" instructions.

Summary updated

Abstract

Science agents increasingly package domain expertise as reusable skills that plan analyses, select tools, and execute workflows. This modularity creates a confidentiality risk: the interactions that deliver a skill's capability can also reveal the skill itself. We build an extensible benchmark of 40 protected skills across eight open- and closed-source science-agent platforms, comparing direct, fixed-strategy, and bandit-guided iterative probing under reference-based and reference-free evaluation. Leakage is pervasive: all eight platforms (100%) leak at least one protected skill, and four reach 100% attack success under a single direct request. To stress-test platforms that resist simple probing, we propose SEEK (Skill Extraction with Evolving Knowledge), a black-box, self-evolving red-teaming agent that consolidates response-derived evidence in persistent memory and evolves reusable attack skills, using only user-visible responses. On 12 skills from SpatialAgent and Biomni, SEEK achieves 100% attack success while reducing the mean first-success round by 31.8% and 38.2% relative to bandit-guided probing. Against a confidentiality policy that blocks all three benchmark protocols, SEEK still extracts one of six protected skills. These findings show that current science agents do not reliably protect their skills and that defenses must be evaluated against adaptive, multi-turn attackers.
BibTeX
@misc{gui2026seek,
  author = {Cheng Gui and Xiang Zheng and Yutao Wu and Hanxun Huang and Yige Li and Kaidi Xu and Cong Wang and Wei-Ying Ma},
  title = {{SEEK: Skill Extraction with Evolving Knowledge against Frontier Science Agents}},
  year = {2026},
  eprint = {2610.34365},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2610.34365},
}
Figure