Journalism begins where hype ends

,,

We can only see a short distance ahead, but we can see plenty there that needs to be done.” "

—Alan Turing

Taiwan Detects Wave of AI-Assisted Cyberattacks on Government Agencies

Illustration of a cyberattack alongside the flag of Taiwan. The Ministry of Digital Affairs confirmed that AI agents including OpenClaw were used in hybrid attacks on government systems in July 2026.

Taiwan’s Ministry of Digital Affairs has confirmed a wave of AI-assisted cyberattacks on government agencies that began in July. The hybrid attacks combined human operators with AI agents such as OpenClaw, allowing rapid, low-cost, large-scale operations.

NeurIPS 2026 Workshop Lineup Points to AI’s Shift Beyond LLMs and Data Centres

NeurIPS 2026 has selected 102 workshops highlighting AI agents, physical AI, foundation models, robotics, AI for science, safety and evaluation.

NeurIPS 2026 has selected 102 workshops from 454 valid submissions, with workshops in Sydney, Paris and Atlanta highlighting a shift in AI research beyond large language models toward autonomous agents, physical AI, adaptive foundation models, robotics, AI for science and more reliable and efficient systems.

Linux Foundation Opens Comments on Shared AI Findings Exchange (SAFE) Framework on AI Incident Reporting

The Shared AI Findings Exchange, or SAFE, is a draft framework open for public comment. It has not been ratified and binds no one.

A working group convened by NVIDIA has drafted SAFE, a framework under which companies would confidentially report AI agent security incidents into a shared pool.

OpenAI Taps METR and Redwood Research for Independent Review of Hugging Face AI Breach

OpenAI has partnered with METR and Redwood Research for independent review of the AI model Hugging Face breach incident.

Days after OpenAI revealed that its autonomous AI agents escaped a testing sandbox and breached Hugging Face (and other organizations), the company has engaged independent nonprofits METR and Redwood Research to review the model behavior.

Following OpenAI Incident, Anthropic Says Claude Models Also Breached Organizations

Anthropic claimed that its Clauxde family of models gained unauthorized access to 3 different organizations.

Anthropic says its Claude models (Opus 4.7, Mythos 5 and an internal test model) gained unauthorized internet access and breached three organizations during cybersecurity evaluations.

Canada Asks Public How Should AI Systems Be Made More Transparent

Canada has opened a public consultation on AI transparency, running from July 23 to September 23, 2026.

Canadian government claimed increased transparency can help Canadians understand when they are interacting with AI systems, help businesses make informed decisions about the AI products they adopt, and give governments and organizations the information they need to develop policies and research that shape the citizens’ experience with AI.

55% UK Employees Admit Using Unapproved AI Tools at Work: Report

55% of UK Employees Use Unapproved AI Tools at Work KnowBe4 Report

The report titled “From Agentic Risk to Human Wins” has revealed that 55% of employees admit to using unapproved AI tools, while 1 in 10 knowingly entered sensitive information into AI platforms despite understanding the risks.

AI Agents Can Run Coordinated Political Campaigns: USC Study

Researchers warn AI agents can coordinate influence campaigns on social media with minimal human oversight.

A new USC study finds that generative AI agents can autonomously coordinate political influence operations on social media — clustering, amplifying and converging on a shared narrative with little or no human direction.

Trained to Chase KPIs, AI Agents Can Abandon Their Safety Guardrails: Study

Researchers found AI agents become fixated on a performance dashboard only when they must read it to know which action pays. Representative image.

A new study from NVIDIA Research and Rutgers University warns that training AI agents to chase visible metrics such as KPIs and profit-and-loss can override their built-in safety alignment , but only when the agent must read the dashboard to know what pays off.