Security
Security at the intersection with AI: adversarial machine learning papers, red-teaming methodologies, defensive ML systems for cybersecurity, and the dual-use research questions that accompany increasingly capable models.
Zscaler Embeds GPT-5.4-Cyber in Zero-Trust Detection Pipeline
Zscaler's TAC membership gives it early access to GPT-5.4-Cyber, embedding the security-tuned frontier model at the core of its detection pipeline and SDLC.
OpenAI Opens GPT-5.4-Cyber to Thousands of Verified Defenders
OpenAI releases GPT-5.4-Cyber with lower refusal boundaries and binary RE capabilities, scaling Trusted Access for Cyber from a limited pilot to thousands of verified security teams.
OpenAI Launches GPT-5.4-Cyber After Anthropic Restricts Mythos
GPT-5.4-Cyber gives defenders a restricted OpenAI model, but independent evaluation remains impossible as both companies compete on AI security framing.
AI Transforms Wireless Networks Through Open RAN
Machine learning is enabling smarter, more efficient cellular networks by tackling spectrum management, resource allocation, and security challenges in open radio access networks—paving the way for 6G.
AI Secures Wireless Networks Against Hidden Attacks
A new AI method dynamically switches between energy-saving and high-performance modes in smart surfaces, while defending against reward poisoning attacks that could cripple network reliability.
A Simpler Way to Measure Confidence in Safety Arguments
Researchers develop a new method to quantify confidence in safety assurance cases, helping engineers balance risk and cost without false precision.
AI Safety Gates Fail, But a Simple Check Succeeds
A new study shows that AI systems cannot reliably self-improve using traditional safety filters, but a verification method achieves perfect safety across scales, including large language models.
Small AI Models Outperform Giants at Predicting Software Bugs
A new ensemble method using compact transformers can detect non-terminating programs more accurately than large language models, offering a practical solution for privacy-sensitive software analysis.
AI Models Learn Tasks, Not Users, for Better Privacy
A new federated learning method trains specialized AI models for specific tasks across distributed data, improving performance by up to 136% when handling multiple or unseen tasks without compromising privacy.
AI Spots Hidden Patterns in Customer Complaints
A new method uses AI to detect sudden drops in user sentiment across social media, helping companies identify real problems before they escalate.
Hidden Safety in AI Models Can Be Revived
Researchers discover that specialized AI models retain safety features but suppress them, and propose a lightweight fix to restore safety without losing performance.
AI Observers Could Make Self-Driving Cars Safer
A new AI layer detects hidden road hazards by understanding context, but a critical flaw in video processing reveals a safety gap that must be fixed before deployment.
AI Generates Fake Network Traffic to Improve Cybersecurity
A new AI method creates synthetic network data that helps intrusion detection systems spot both known and unknown attacks more accurately, boosting detection rates by up to 9.3% in tests.
Quantum Computers Could Crack Bitcoin Security Sooner Than Expected
New research reveals that quantum computers may break the cryptography securing billions in cryptocurrencies within years, forcing urgent upgrades to protect digital assets.
New Tool Tests AI Agent Skills for Safety and Usefulness
SkillTester evaluates third-party AI skills by comparing them to baseline performance and running security probes, helping users avoid malicious or ineffective tools in growing agent ecosystems.
AI Safety Breakthrough for Small Language Models
A new defense method can detect and block malicious prompts in real-time without slowing down AI responses, making small language models safer for everyday use.
AI Backdoor Attacks Hide in Compressed Data
A new stealthy attack method can embed hidden triggers into condensed datasets, compromising AI models without detection while maintaining normal performance on clean tasks.
AI Safety Tests Miss How Models Amplify Human Harm
Researchers propose measuring 'harmful capability uplift'—how much AI increases users' ability to cause damage—arguing current safety evaluations fail to capture real-world risks where humans and AI collaborate on malicious tasks.
New AI Safety System Prevents Digital Assistants from Going Rogue
Researchers developed a framework that stops AI agents from executing harmful actions like deleting emails or leaking data, achieving over 95% accuracy in real-world tests.
AI Safety Gains a Database Theory Upgrade
A new equivalence shows that checking AI agent safety is exactly the same as evaluating a simple database query, unlocking decades of algorithmic results for faster and more reliable systems.
AI Could Make Online Dating Safer by Reading Nonverbal Cues
A new research agenda proposes using computer vision to detect discomfort and disinterest in video dates, aiming to close a communication gap that disproportionately harms women and vulnerable users.
AI Cybersecurity Tools Face Hidden Limits
A new mathematical theory reveals that AI cannot boost security operations past human bottlenecks, and common assumptions about false alarms are flawed.
New AI Method Protects Distributed Training from Attacks
A novel approach combines redundant computation and encoding to defend against malicious devices in distributed learning, reducing solution error and improving communication efficiency.
New Hardware Security Method Defeats AI Attacks
A simple resistor-capacitor device can protect Internet of Things gadgets from machine learning-based hacking attempts, offering a lightweight alternative to complex encryption.