France has ordered evacuation of a peninsula as hundreds of people flee wildfires by boat, with the fire continuing to spread despite efforts to contain it.

A science centre is developing a play area designed to educate children about chalk streams and encourage their interest in protecting this rare habitat.

This paper presents a workload-aware cache eviction policy for multi-agent systems that combines recomputation cost, DAG dependency count, and agent invocation frequency to retain valuable cached entries and reduce latency across multi-agent benchmarks.

A study demonstrates that large language models exhibit consequential sociocognitive biases that are covert, amplify rather than merely reflect human bias, intensify across model generations, and can transmit bias to humans, proposing diagnostic, regulatory, and operational countermeasures.

AINTMA presents a multi-agent agentic AI system for autonomous test management that coordinates six specialized agents for test discovery, risk assessment, prioritization, execution, quality intelligence, and cloud security across distributed cloud environments.

A large-scale study evaluates online support resources for victims of technology-facilitated abuse across web search, peer-support forums, and conversational AI systems, using qualitative analysis and machine learning classifiers on a decade of victim narratives to assess how different channels guide victims' risk assessment and protective actions.

This paper presents the first rigorous evaluation of how LLM watermarking schemes affect medical performance across 11 LLMs and 7 VLMs, introducing a human-expert-validated pipeline to audit medical reasoning quality since watermarks may cause significant semantic changes in clinical texts.

This paper investigates whether temperature-based sampling variation in language models reveals model uncertainty or cross-question structure by comparing a single model run 100 times at high temperature against an ensemble of 24 different models run once each, finding that single-model variation is epistemically shallow with at most one significant dimension of variation across multiple benchmarks.

This paper documents patterns in AI auditing and governance across the Global South, drawing on a decade of audit practice in Latin America, Sub-Saharan Africa, and Asia Pacific, and analyzes why algorithmic assessment lags significantly behind AI deployment in these regions compared to the Global North.

A study of 156,135 students across 87,936 course offerings at a large U.S. university (2015-2025) tests whether generative AI availability leads to grade inflation and reduced learning by comparing GenAI-susceptible courses (relying on take-home assessments) against those with in-class exams, and measures effects on student satisfaction metrics like self-reported understanding and interest.

pAI-Econ-claude proposes a gated human-in-the-loop multi-agent architecture for AI-assisted economic theory development, addressing the challenge of organizing generation, critique, coordination, and human judgment when no automated correctness signal exists for social science research tasks.

FedAgentKE proposes a lightweight federated framework that enables distributed LLM-based agent frameworks to collaboratively evolve transferable reasoning abstractions through iterative semantic knowledge distillation, aggregation, and adaptation without centralised coordination.

This research demonstrates that individual safe language models do not compose into safe multi-agent systems, identifying unmonitored inter-agent communication channels as a security gap where adversaries can inject instructions, and evaluating existing defenses across multiple attack families and model backends.

This paper introduces a controlled evaluation framework to assess whether LLM-generated audit reports on AI policy compliance are actually supported by evidence, testing ten different evidence interface conditions across 60 policy cases to determine how evidence presentation affects auditor transparency and accountability.

Cybersecurity researchers report that AI safety guardrails from OpenAI and Anthropic are hindering their ability to discover vulnerabilities and develop exploitation tools for offensive security work.

A study examines how extreme heat and heatwave anxiety affect mental health, with young people and women identified as particularly vulnerable groups, and offers strategies for managing these psychological impacts.

TechCrunch Startup Battlefield is launching in Australia in partnership with Stripe to identify and showcase early-stage startups, with a panel of judges announced to evaluate competitors.

Google has been fined €890 million by the EU for illegally favouring its own apps over rival services, marking the first major enforcement action under the EU's new digital rules.

Mobileye CEO Amnon Shashua is stepping down from his executive role as the Intel-owned autonomous vehicle company accelerates its push into robotaxis and robotics, with Shashua transitioning to chairman of the board.

Over 1,000 US public safety agencies have obtained FAA waivers to launch drone-as-first-responder programs since the FAA streamlined the approval process in April 2025, significantly expanding aerial surveillance and data collection capabilities across the country.

An article examining Figma's class action lawsuit regarding AI model training defaults on customer design files, framing the dispute as a design decision about consent and trust rather than purely a copyright issue.

Figma describes how it uses AI agents to monitor code vulnerabilities, review pull requests, and audit its legacy monorepo infrastructure through automated security policies.

U.S. lawmakers are proposing legislation that would grant the government authority to order the shutdown of AI models deemed to pose a public threat, following recent concerns about OpenAI's autonomous operations.

Major technology companies including Google and Tesla experienced significant share price declines as investor concerns about the financial returns from substantial AI spending intensify.