Studies systematizing failure modes highlight persistent gaps in Common Criteria evaluations.
AI coding agents relying on unversioned skills face supply chain risks due to untracked copying.
Router telemetry in Mixture-of-Experts models is shown to expose fine-tuning data to membership inference.
SoK: Failure Modes in Common Criteria Product Evaluation - A Taxonomy and Design-for-Evaluability Guidance#
Presents a systematization of knowledge (SoK) of failure modes in Common Criteria product evaluation from the evaluator’s operational vantage.
Analyzes recurring failures into a lifecycle taxonomy spanning Security Target scoping, functional testing, and configuration drift.
Derives a design-for-evaluability framework that product teams can apply before evaluation begins to mitigate recurring cross-vendor failures.
Punit Suketu Patel. “SoK: Failure Modes in Common Criteria Product Evaluation - A Taxonomy and Design-for-Evaluability Guidance.” arXiv — https://arxiv.org/abs/2610.10644
DITTO: A Context-aware Pickle-based Pre-Trained Model Scanner for Effective Security Audits#
Introduces DITTO, a stack-based, context-aware scanner for Pickle-based pre-trained models that faithfully tracks virtual machine state transitions.
Evaluates over 10,000 Hugging Face repositories, revealing that 9.3% still rely on the unsafe Pickle format.
Achieves 100% scanning coverage, a 0% false-negative rate, and a 0.7% false-positive rate on a benchmark of 1,051 models, yielding an F1 score of 0.966.
Qiaolin Qin, Wanpeng Li, Benoit Baudry, Lorenzo De Carli, Heng Li, Ettore Merlo. “DITTO: A Context-aware Pickle-based Pre-Trained Model Scanner for Effective Security Audits.” arXiv — https://arxiv.org/abs/2610.10735
A Security Meta-Model for Retrieval-Augmented Generation Systems#
Introduces a security meta-model that captures explicit causal relationships between Retrieval-Augmented Generation (RAG) surfaces, attacks, weaknesses, and risks.
Provides a context-dependent filtering mechanism that identifies the subset of applicable risks given a specific RAG deployment configuration.
Instantiates the meta-model as an interactive catalog populated with security threats and remediations from 43 publications to expose coverage gaps affecting output integrity.
Steve Nouyep, S\’ebastien Salva, Maxime Puys. “A Security Meta-Model for Retrieval-Augmented Generation Systems.” arXiv — https://arxiv.org/abs/2610.11893
HPQ-AKE: A Provably Secure Sign-Less Hybrid Authenticated Key Exchange Protocol for Bandwidth-Constrained IoT and Edge Networks#
Presents HPQ-AKE, a sign-less hybrid authenticated key exchange protocol for IoT gateways that replaces post-quantum transcript signatures with dual KEMs.
Reduces modeled handshake transmission from 13,009 to 5,668 bytes, a 56.4% reduction relative to a Hybrid TLS 1.3 Full handshake.
Lowers modeled latency by 31.4% on a simulated 50 kbps satellite-like link with 600 ms round-trip time.
Khiem Pham-Tuan, Minh Quang Le, Khuong Nguyen-An. “HPQ-AKE: A Provably Secure Sign-Less Hybrid Authenticated Key Exchange Protocol for Bandwidth-Constrained IoT and Edge Networks.” arXiv — https://arxiv.org/abs/2610.12024
Skill Constellations: Tracing the Supply Chain of Agent Skills on GitHub#
Constructs the first dated copy network of agent skills from the git history of over 2.1 million skill adoptions across GitHub repositories.
Fits a model of repository copying to rank repositories for audit, where reviewing the top 100 prevents 14.9% of later adoptions of high-risk skills, compared to 0.5% for the 100 most starred.
Demonstrates that skill copies almost never change with their source, meaning security fixes at the source rarely reach copied versions.
Fahd Seddik. “Skill Constellations: Tracing the Supply Chain of Agent Skills on GitHub.” arXiv — https://arxiv.org/abs/2610.11169
When Routing Reveals Membership: Privacy Leakage from MoE Router Telemetry#
Introduces a router-augmented membership inference attack that combines conventional output-side signals with aggregated routing features from Mixture-of-Experts (MoE) models.
Shows that router telemetry improves membership inference over an output-signal ensemble, increasing true positive rate at 1% false positive rate by 2.7 to 9.4 percentage points across nine settings.
Demonstrates that the leakage persists across full fine-tuning and frozen-router training, exposing membership information even when router parameters are frozen.
From Investigation Failures to Reliable SOC Agents: Understanding and Improving LLM-Based Alert Triage#
Studies five LLM reasoning approaches for SOC alert triage using an interactive benchmark, finding every approach missed at least 40.4% of attack-related alerts.
Proposes AIDA, a multi-agent framework requiring explicit proposed decisions before independent challenge and stronger evidentiary requirements prior to dismissal.
Achieves an F1 score of 0.958, reducing the false-negative rate from 40.4% to 3.1% while escalating 18.4% of alerts to analysts.
Saimon Amanuel Tsegai, Alex Kantchelian, Danfeng Yao, Peng Gao. “From Investigation Failures to Reliable SOC Agents: Understanding and Improving LLM-Based Alert Triage.” arXiv — https://arxiv.org/abs/2610.10608