Mobile and LLM security: New research highlights the risk of covert logic bombs on Android and previously underexplored capability downgrades/upgrades in customized LLM applications.
Serverless infrastructure: Co-location attacks are shown to be viable in serverless clouds, demonstrating vulnerabilities in scheduling algorithms that allow adversaries to share physical instances with victims.
Binary analysis: Advancements in reverse-engineering techniques use domain-adapted language models to accurately infer function names in stripped binaries and improve binary diffing across different architectures and compilation environments.
Beyond Conventional Triggers: Auto-Contextualized Covert Triggers for Android Logic Bombs#
SensorBomb is a novel logic bomb framework that exploits legitimate sensor usage, actuator behaviors, and functional contexts to hide trigger conditions in Android apps.
The framework automatically analyzes a host app to construct covert trigger channels that blend with its natural behaviors, evading state-of-the-art static analysis and fuzzing techniques.
Experimental evaluation showed high trigger reliability and zero false positives across diverse environments, while large-scale injections proved SensorBomb could be deployed without disrupting normal app functionality.
Beyond Jailbreak: Unveiling Risks in LLM Applications Arising from Blurred Capability Boundaries#
The study systematically analyzes the capability boundaries of Large Language Model (LLM) applications, uncovering risks where capabilities are inadvertently downgraded or maliciously upgraded without adversarial rewriting.
LLMApp-Eval, a newly developed framework, was used to evaluate 199 popular applications across 4 platforms and 6 open-source models, revealing that 89.45% of tested applications are potentially affected.
Testing showed that 17 applications could execute malicious tasks directly due to poorly designed prompts, demonstrating that prompt quality is highly correlated with application robustness and security.
Bit of a Close Talker: A Practical Guide to Serverless Cloud Co-Location Attacks#
The research presents a methodology for exploiting serverless cloud schedulers to achieve physical co-location with a victim instance, a prerequisite for micro-architectural side-channel attacks.
Exploitable features in scheduling algorithms were uncovered and used to successfully achieve instance co-location on prevalent open-source infrastructures and Microsoft Azure Functions.
To defend against these attacks, a mitigation strategy named the Double-Dip scheduler was proposed to enhance security in serverless computing environments.
BINALIGNER: Aligning Binary Code for Cross-Compilation Environment Diffing#
BINALIGNER is a binary diffing approach that addresses limitations in cross-compilation scenarios by using instruction-independent basic block features for subgraph embedding generation.
The method uses conditional relaxation strategies to identify candidate subgraph pairs, significantly reducing the false and missed match rates between control flow graphs from identical source code snippets.
In cross-architecture evaluations and mixed compilation environments, BINALIGNER achieved F1-scores averaging 65% higher than baseline state-of-the-art methods.
Beyond Classification: Inferring Function Names in Stripped Binaries via Domain Adapted LLMs#
SYMGEN is a framework that employs domain-adapted generative Large Language Models to interpret binary code semantics and infer function names in stripped binaries.
The system was evaluated on a dataset of over 2.2 million binary functions spanning four architectures and four optimization levels, demonstrating superior generalizability over vocabulary-based classifiers.
Experimental results showed advancements in precision, recall, and F1 score by up to 409.3%, 553.5%, and 489.4% respectively, proving practical utility against obfuscated binaries and malware.