Searcharxiv⌕ Search
Home/Archive/arXiv

arXiv · 2609.13353

SkillAtlas: An Attack Trace Library for Agent Skills

Yuxin Tian·Zenghao Duan·Liang Pang·Zhiyi Yin·Xueqi Cheng
View original source ↗Read source document ↗Export metadata ↓

Abstract

Agent skills are reusable units for language-model agents, but their risks emerge through model decisions, user context, tool calls, and execution feedback rather than through stable signatures or a single sandbox run. Existing static, dynamic, and benchmark-style evaluations rarely preserve public evidence that can be inspected, searched, and reused. We present SkillAtlas, a hosted attack trace library that converts private agent-skill security report bundles into reviewed, redacted, and searchable public cases. The library contains 3,014 cases, 6,589 traces, 151,131 steps, 233 affected skills, and 8 risk categories; 42.5% of successful cases first become successful after a non-success initial round, and trajectory-grounded labels improve pre-execution guard accuracy to 0.770.

Explore related subjects

cs.CRcs.AI

Keep this discovery

Explore connections, maps & timelines

BibTeXRIS

Yuxin Tian, Zenghao Duan, Liang Pang, Zhiyi Yin, Xueqi Cheng. 2026-09-11. SkillAtlas: An Attack Trace Library for Agent Skills. https://arxiv.org/abs/2609.13353

Cite the original work for its findings. Save a collection to share your selection of sources.

KEEP EXPLORING

Related papers

arXiv2026

SynGhost: Invisible and Universal Task-agnostic Backdoor Attack via Syntactic Transfer

Although pre-training achieves remarkable performance, it suffers from task-agnostic backdoor attacks due to vulnerabilities in data and training mechanisms. These attacks can transfer backdoors to various downstream tasks. In this paper, we introduce $\mathtt{maxEntropy}$, an entropy-based poisoning filter that mitigates such risks. To overcome the limitations of manual target setting and explicit triggers, we propose $\mathtt{SynGhost}$, an invisible and universal task-agnostic backdoor attack via syntactic transfer, further exposing vulnerabilities in pre-trained language models (PLMs). Specifically, $\mathtt{SynGhost}$ injects multiple syntactic backdoors into the pre-training space through corpus poisoning, while preserving the PLM's pre-training capabilities. Second, $\mathtt{SynGhost}$ adaptively selects optimal targets based on contrastive learning, creating a uniform distribution in the pre-training space. To identify syntactic differences, we also introduce an awareness module to minimize interference between backdoors. Experiments show that $\mathtt{SynGhost}$ poses significant threats and can transfer to various downstream tasks. Furthermore, $\mathtt{SynGhost}$ resists defenses based on perplexity, fine-pruning, and $\mathtt{maxEntropy}$. The code is available at https://github.com/Zhou-CyberSecurity-AI/SynGhost.

cs.CR↗
arXiv2026

Towards the ideals of Self-Recovery and Metadata Privacy in Social Vault Recovery with Apollo

Social recovery enables users to enlist trusted contacts, or trustees, to help recover lost access to end-to-end-encrypted repositories or vaults. However, existing recovery mechanisms often make strong memorability assumptions about what users will remember. Weakening these memorability assumptions to increase the robustness of recovery is possible, but may leak sensitive metadata about the user and/or trustees if done naively. This paper's first contribution is to draw attention to and formalize this basic tension between memorability and metadata privacy in social vault recovery. Our second contribution is Apollo, a social recovery mechanism that aims to avoid any memorability assumptions while strongly protecting recovery metadata privacy. Apollo approximates the ideal of self-recovery by relying only on a threshold of social reconnection events, which may be initiated either by the user or the user's contacts. Apollo thereby has a chance of succeeding in vault recovery even in a worst-case scenario where the user has forgotten all metadata, including even the vault's existence. To protect the metadata's privacy, Apollo distributes either real or fake (chaff) data to all of a user's contacts, not just the user's trustees, thus systematically anonymizing the trustees among the larger set of contacts. To make this anonymity set scalable, Apollo uses a novel multi-layered secret sharing scheme to mitigate the computational overhead of recovery in this setting, which would otherwise be exponential in the recovery threshold. Finally, we evaluate a prototype implementation of Apollo. Apollo reduces the probability of malicious recovery to under 0.1% for an adversary capable of obtaining shares from every 1-out-of-2 contacts. After reconnecting with 30 contacts, the multi-layered design shows an improvement of 5 orders in computation time, compared to a single-layered approach.

cs.CR↗
arXiv2026

Computational Certified Deletion Property of Magic Square Game and its Application to Classical Secure Key Leasing

We present the first construction of a computational Certified Deletion Property (CDP) achievable with classical communication, derived from the compilation of the non-local Magic Square Game (MSG). We leverage the KLVY compiler to transform the non-local MSG into a 2-round interactive protocol, rigorously demonstrating that this compilation preserves the game-specific CDP. Previously, the quantum value and rigidity of the compiled game were investigated. We emphasize that we are the first to investigate CDP (local randomness in [Fu and Miller, Phys. Rev. A 97, 032324 (2018)]) for the compiled game. Then, we combine this CDP with the framework [Kitagawa, Morimae, and Yamakawa, Eurocrypt 2025] to construct Secure Key Leasing with classical Lessor (cSKL). SKL enables the Lessor to lease the secret key to the Lessee and verify that a quantum Lessee has indeed deleted the key. In this paper, we realize cSKL for PKE, PRF, and digital signature. Compared to prior works for cSKL, we realize cSKL for PRF and digital signature for the first time. In addition, we succeed in weakening the assumption needed to construct cSKL.

cs.CR↗