When Good Sounds Go Adversarial: Jailbreaking Audio-Language Models with Benign Inputscs.AI updates on arXiv.org arXiv:2508.03365v3 Announce Type: replace-cross
Abstract: As large language models (LLMs) become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that manipulates state-of-the-art audio language models to generate harmful content. Our method embeds harmful payloads as subtle perturbations into audio inputs that remain intelligible to human listeners. The first stage uses a novel reward-based white-box optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to jailbreak the target model and elicit harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use gradient-based optimization to embed subtle perturbations into benign audio carriers, such as weather queries or greeting messages. Our method achieves average attack success rates of 60-78% across two benchmarks and five multimodal LLMs, validated by multiple evaluation frameworks. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating multimodal AI systems.
arXiv:2508.03365v3 Announce Type: replace-cross
Abstract: As large language models (LLMs) become increasingly integrated into daily life, audio has emerged as a key interface for human-AI interaction. However, this convenience also introduces new vulnerabilities, making audio a potential attack surface for adversaries. Our research introduces WhisperInject, a two-stage adversarial audio attack framework that manipulates state-of-the-art audio language models to generate harmful content. Our method embeds harmful payloads as subtle perturbations into audio inputs that remain intelligible to human listeners. The first stage uses a novel reward-based white-box optimization method, Reinforcement Learning with Projected Gradient Descent (RL-PGD), to jailbreak the target model and elicit harmful native responses. This native harmful response then serves as the target for Stage 2, Payload Injection, where we use gradient-based optimization to embed subtle perturbations into benign audio carriers, such as weather queries or greeting messages. Our method achieves average attack success rates of 60-78% across two benchmarks and five multimodal LLMs, validated by multiple evaluation frameworks. Our work demonstrates a new class of practical, audio-native threats, moving beyond theoretical exploits to reveal a feasible and covert method for manipulating multimodal AI systems. Read More
Intuit, Uber, and State Farm trial AI agents inside enterprise workflowsAI News The way large companies use artificial intelligence is changing. For years, AI in business meant experimenting with tools that could answer questions or help with small tasks. Now, some big enterprises are moving beyond tools to AI agents that can actually do practical work in systems and workflows. This week, OpenAI introduced a new platform
The post Intuit, Uber, and State Farm trial AI agents inside enterprise workflows appeared first on AI News.
The way large companies use artificial intelligence is changing. For years, AI in business meant experimenting with tools that could answer questions or help with small tasks. Now, some big enterprises are moving beyond tools to AI agents that can actually do practical work in systems and workflows. This week, OpenAI introduced a new platform
The post Intuit, Uber, and State Farm trial AI agents inside enterprise workflows appeared first on AI News. Read More
How separating logic and search boosts AI agent scalabilityAI News Separating logic from inference improves AI agent scalability by decoupling core workflows from execution strategies. The transition from generative AI prototypes to production-grade agents introduces a specific engineering hurdle: reliability. LLMs are stochastic by nature. A prompt that works once may fail on the second attempt. To mitigate this, development teams often wrap core business
The post How separating logic and search boosts AI agent scalability appeared first on AI News.
Separating logic from inference improves AI agent scalability by decoupling core workflows from execution strategies. The transition from generative AI prototypes to production-grade agents introduces a specific engineering hurdle: reliability. LLMs are stochastic by nature. A prompt that works once may fail on the second attempt. To mitigate this, development teams often wrap core business
The post How separating logic and search boosts AI agent scalability appeared first on AI News. Read More
Prompt Fidelity: Measuring How Much of Your Intent an AI Agent Actually ExecutesTowards Data Science
Prompt Fidelity: Measuring How Much of Your Intent an AI Agent Actually ExecutesTowards Data Science How much of your AI agent’s output is real data versus confident guesswork?
The post Prompt Fidelity: Measuring How Much of Your Intent an AI Agent Actually Executes appeared first on Towards Data Science.
How much of your AI agent’s output is real data versus confident guesswork?
The post Prompt Fidelity: Measuring How Much of Your Intent an AI Agent Actually Executes appeared first on Towards Data Science. Read More
The U.S. Cybersecurity and Infrastructure Security Agency (CISA) has issued a new binding operational directive requiring federal agencies to identify and remove network edge devices that no longer receive security updates from manufacturers. […] Read More
Cybersecurity researchers have discovered a new supply chain attack in which legitimate packages on npm and the Python Package Index (PyPI) repository have been compromised to push malicious versions to facilitate wallet credential theft and remote code execution. The compromised versions of the two packages are listed below – @dydxprotocol/v4-client-js (npm) – 3.4.1, 1.22.1, 1.15.2, […]
GPT-5 lowers the cost of cell-free protein synthesisOpenAI News An autonomous lab combining OpenAI’s GPT-5 with Ginkgo Bioworks’ cloud automation cut cell-free protein synthesis costs by 40% through closed-loop experimentation.
An autonomous lab combining OpenAI’s GPT-5 with Ginkgo Bioworks’ cloud automation cut cell-free protein synthesis costs by 40% through closed-loop experimentation. Read More
AI Expo 2026 Day 2: Moving experimental pilots to AI productionAI News The second day of the co-located AI & Big Data Expo and Digital Transformation Week in London showed a market in a clear transition. Early excitement over generative models is fading. Enterprise leaders now face the friction of fitting these tools into current stacks. Day two sessions focused less on large language models and more
The post AI Expo 2026 Day 2: Moving experimental pilots to AI production appeared first on AI News.
The second day of the co-located AI & Big Data Expo and Digital Transformation Week in London showed a market in a clear transition. Early excitement over generative models is fading. Enterprise leaders now face the friction of fitting these tools into current stacks. Day two sessions focused less on large language models and more
The post AI Expo 2026 Day 2: Moving experimental pilots to AI production appeared first on AI News. Read More
How to Build Production-Grade Data Validation Pipelines Using Pandera, Typed Schemas, and Composable DataFrame ContractsMarkTechPost Schemas, and Composable DataFrame ContractsIn this tutorial, we demonstrate how to build robust, production-grade data validation pipelines using Pandera with typed DataFrame models. We start by simulating realistic, imperfect transactional data and progressively enforce strict schema constraints, column-level rules, and cross-column business logic using declarative checks. We show how lazy validation helps us surface multiple
The post How to Build Production-Grade Data Validation Pipelines Using Pandera, Typed Schemas, and Composable DataFrame Contracts appeared first on MarkTechPost.
Schemas, and Composable DataFrame ContractsIn this tutorial, we demonstrate how to build robust, production-grade data validation pipelines using Pandera with typed DataFrame models. We start by simulating realistic, imperfect transactional data and progressively enforce strict schema constraints, column-level rules, and cross-column business logic using declarative checks. We show how lazy validation helps us surface multiple
The post How to Build Production-Grade Data Validation Pipelines Using Pandera, Typed Schemas, and Composable DataFrame Contracts appeared first on MarkTechPost. Read More
OpenAI Just Launched GPT-5.3-Codex: A Faster Agentic Coding Model Unifying Frontier Code Performance And Professional Reasoning Into One SystemMarkTechPost OpenAI has just introduced GPT-5.3-Codex, a new agentic coding model that extends Codex from writing and reviewing code to handling a broad range of work on a computer. The model combines the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2 into a single system, and it runs 25% faster
The post OpenAI Just Launched GPT-5.3-Codex: A Faster Agentic Coding Model Unifying Frontier Code Performance And Professional Reasoning Into One System appeared first on MarkTechPost.
OpenAI has just introduced GPT-5.3-Codex, a new agentic coding model that extends Codex from writing and reviewing code to handling a broad range of work on a computer. The model combines the frontier coding performance of GPT-5.2-Codex with the reasoning and professional knowledge capabilities of GPT-5.2 into a single system, and it runs 25% faster
The post OpenAI Just Launched GPT-5.3-Codex: A Faster Agentic Coding Model Unifying Frontier Code Performance And Professional Reasoning Into One System appeared first on MarkTechPost. Read More
