Federal safety officials are investigating a deadly cargo jet crash that involved AI technology, raising serious questions about the risks of relying on automation for critical safety functions. The probe highlights the potential consequences when AI systems fail in high-stakes environments where human lives are at stake.
Meanwhile, Meta has developed a thin orchestrator that significantly improves the reliability of AI agents. The system, which consists of just 1,300 lines of code, routes 20 workflows to 14 specialized agents and focuses entirely on reliability checks rather than domain-specific work. A study at Mount Sinai found this approach achieved 65.3% accuracy compared to just 16.6% for single agents, while reducing token usage by up to 65-fold.
On the threat side, AI is accelerating ransomware attacks on backup systems. Research shows 94% of ransomware victims faced attempts to compromise backups, with 57% of those attacks succeeding. Attackers now reach backup systems in hours rather than weeks, and one attacker remained undetected for five weeks before striking, according to a Check Point report.
Military documents have revealed a plan for an agentic AI tool called ECHOTRIBBLE, designed to hack adversaries. The tool remains in the whitepaper phase, but Army Central Command is seeking private sector help. Experts warn that autonomous hacking AI systems pose serious dangers given known failure modes, especially as Secretary of Defense Pete Hegseth has pushed for military AI experimentation.
In the academic world, AI firms have been solving previously unsolvable mathematical problems, shaking the research community. OpenAI released solutions to 719 previously unsolved problems, prompting Fields Medal winner Hugo Duminil-Copin to say all his avenues of research are now gone. The Simons Foundation announced a $15 million program co-funded by OpenAI and Anthropic to support mathematicians navigating this shift. Experts stress that math remains central as a tool for clear human thinking.
On the policy front, the Maltese government plans to introduce AI training across all 36 Active Ageing Centres. The AI4ALL scheme, backed by an agreement between the Malta Digital Innovation Authority and Active Ageing and Community Care, aims to strengthen older persons' independence by teaching them to use AI safely. Staff will be trained to guide elderly participants through the process.
Security experts warn that fragmented AI security is creating dangerous blind spots for enterprises. Organizations increasingly use multiple generative AI models and agents, but few govern access to core business systems effectively. Individual model-level controls can miss threats that emerge when multiple approved components interact, and attackers are already using generative AI for reconnaissance, phishing, and deepfake fraud.
Key Takeaways
- Federal officials are investigating a deadly cargo jet crash involving AI technology, probing the risks of automation in critical safety roles
- Meta's thin orchestrator achieved 65.3% accuracy versus 16.6% for single agents in a Mount Sinai study, while cutting token usage by up to 65-fold
- AI has sped up ransomware attacks on backups, with 94% of victims facing compromise attempts and 57% succeeding
- Military documents reveal a plan for ECHOTRIBBLE, an agentic AI tool aimed at hacking adversaries, still in whitepaper phase
- OpenAI released solutions to 719 previously unsolved math problems, prompting Fields Medal winner Hugo Duminil-Copin to say his research avenues are gone
- The Simons Foundation announced a $15 million program co-funded by OpenAI and Anthropic to support mathematicians
- Malta will roll out AI training across all 36 Active Ageing Centres under the AI4ALL scheme to boost elderly independence
- Enterprises face growing security blind spots as fragmented AI tools gain access to core business systems without effective governance
- Attackers are using generative AI for reconnaissance, phishing, and deepfake fraud, while models' refusal behaviors can weaken without retraining
- Science journalists are urged to understand the differences between discriminative and generative AI as LLMs like GPT and Claude spread across research fields
What happens when good AI goes bad
This report examines what occurs when AI systems fail or cause harm. Federal safety officials are investigating a deadly cargo jet crash involving AI technology. The story explores the risks of relying on AI for critical safety functions. Reporter Joce Sterman and video producer Scotty Smith contributed to the investigation.
Federal officials probe deadly cargo jet crash and AI safety technology
Federal safety officials are investigating a deadly cargo jet crash. The report looks at technology designed to save lives. Reporter Joce Sterman and video producer Scotty Smith worked on the story.
Meta's thin orchestrator makes AI agents more reliable and efficient
Meta built a 1,300-line orchestrator that routes 20 workflows to 14 specialized AI agents. The orchestrator does zero domain work and focuses on reliability checks. A study at Mount Sinai found this approach held 65.3% accuracy compared to 16.6% for single agents. The team reduced token usage up to 65-fold by splitting tasks between a router and specialist agents.
AI speeds up ransomware attacks on backups and storage systems
Backups are the first target in ransomware attacks, and AI now lets attackers reach them in hours instead of weeks. Research found 94% of ransomware victims saw attackers try to compromise backups, and 57% of those attempts succeeded. A Check Point report showed an attacker stayed undetected for five weeks before striking. Experts say hardening must be continuous because attackers move faster every year.
Military documents reveal plan for AI that hacks adversaries
Documents reveal a military plan for an agentic AI tool called ECHOTRIBBLE. The tool is still in the whitepaper phase and aims to hack U.S. adversaries. Army Central Command is seeking private sector help. Secretary of Defense Pete Hegseth previously called for military AI experimentation. Experts warn that autonomous hacking AI systems are incredibly dangerous given known failure modes.
Elderly residents to receive AI training across all 36 Active Ageing Centres
The Maltese government announced that AI training will be introduced in all 36 Active Ageing Centres. The AI4ALL scheme will be extended to elderly persons' centres and residential care homes. An agreement was signed between the Malta Digital Innovation Authority and Active Ageing and Community Care. Staff will receive training to guide elderly participants in using AI safely. Deputy Prime Minister Ian Borg said the initiative aims to strengthen older persons' independence.
How AI is changing science and what journalists need to understand
Researchers are rapidly adopting AI for data analysis and entire research studies. Science journalists need to understand the basic principles and differences between discriminative AI and generative AI. Large language models like GPT and Claude are being used across nearly every scientific field. Mathematicians report that LLM-based AI agents have broken entire fields of research. Knowing about different kinds of AI helps journalists frame findings appropriately for readers.
AI models crack previously unsolvable math problems challenging researchers
AI firms have been publishing solutions to previously unsolved mathematical problems, shaking top-level mathematicians. OpenAI recently released solutions to 719 previously unsolved problems. Fields Medal winner Hugo Duminil-Copin said all his avenues of research are now gone. The Simons Foundation announced a $15 million program co-funded by OpenAI and Anthropic to support mathematicians. Experts stress that math remains central as a tool for clear human thinking.
Fragmented AI security creating dangerous blind spots for enterprises
Enterprises are adopting a growing mix of generative AI models and agents, increasing the cyber-attack surface. Many organizations say AI tools have access to core business systems but few govern that access effectively. Individual model-level controls may miss threats that emerge when multiple approved components act together. Attackers are already using generative AI for reconnaissance, phishing, and deepfake fraud. A model's refusal behavior can weaken without retraining, making external governance checks essential.
A practical glossary explaining key AI terms and concepts
This glossary covers important AI concepts from capability to chatbots. Capability means what an AI system can do, but not necessarily how reliably it does it. Catastrophic forgetting occurs when a model learns new information but loses earlier knowledge. Causal AI focuses on cause and effect rather than just correlations. Chain-of-thought refers to step-by-step reasoning that may help models solve problems. ChatGPT is an AI assistant using large language models for tasks like writing and research support.
Sources
- What happens when good AI goes bad?
- What happens when good AI goes bad?
- The Thin Orchestrator: How We Made Our AI Agents More Reliable by Giving the Router Less to Do
- Hardening Is Continuous: Protecting Storage and Backup in the AI Era
- Documents Reveal Military Plan for Agentic AI That Hacks U.S. Adversaries
- AI for all ages: Elderly home residents will get training in use of artificial intelligence
- AI is Changing Science: Here’s What Journalists Need to Know
- AI challenges math by solving the unsolvable
- The hidden risks of fragmented AI security
- Artificial Intelligence: Part 5
Comments
Please log in to post a comment.