Microsoft Updates AI Threat Taxonomy
In brief
- Microsoft updated its AI threat taxonomy to include seven new failure mode categories.
- The update is based on 12 months of red team engagements against deployed agentic systems.
- The new taxonomy includes failure modes such as agentic supply chain compromise and tool poisoning.
- Open-source agentic frameworks like OpenClaw have accumulated over 336,000 GitHub stars and spawned more than 2,100 agents.
- These frameworks have introduced new vulnerabilities, including 512 identified in OpenClaw.
- The updated taxonomy will help security teams better understand and address AI-related threats.
- New threats will continue to emerge as AI technology advances.
Terms in this brief
- AI threat taxonomy
- A classification system used to identify and categorize potential threats related to AI technologies. This helps security teams understand risks and develop strategies to mitigate them.
- red team engagements
- Simulated attacks conducted by cybersecurity professionals (red teams) to test the robustness of AI systems against real-world threats, helping to uncover vulnerabilities before they can be exploited.
Read full story at Microsoft →
More briefs
OpenAI Files for Public Stock Offering
OpenAI has filed to offer its stock on public markets. The company made this announcement just a week after its rival Anthropic did the same. The move is important because OpenAI's valuation is $852 billion. The company has hundreds of millions of downloads for its ChatGPT app. OpenAI will use the public offering to gain more funds for its AI models and data centers. The company will soon go public.
NVIDIA Enhances AI Collaboration Through Federated Learning Breakthroughs
NVIDIA has introduced a new approach in federated learning, enabling AI models to collaborate more effectively without sharing sensitive data. This advancement allows institutions to work together on tasks like genomics research, improving patient outcomes while maintaining privacy. By using innovative aggregation techniques, the method addresses challenges that previously hindered progress in fields such as healthcare and finance. This development is significant because it balances the need for collaborative AI advancements with strict data security requirements. Federated learning typically struggles with varying participant resources and communication efficiency, but NVIDIA's solution streamlines these processes, making large-scale collaborations more feasible. For instance, researchers can now train models on decentralized data from multiple sources without compromising performance or privacy. Looking ahead, this breakthrough could pave the way for more secure and efficient AI research across industries. Developers and researchers should watch for further refinements in how these techniques are applied to real-world problems, potentially leading to faster discoveries and better decision-making tools.
AI Showcases Strong Potential for Automating Data Extraction from Dutch Neuroradiology Reports
AI has demonstrated impressive ability to extract data from complex medical reports, a breakthrough that could transform how radiologists handle their work. In a recent study, researchers tested the LLaMA 3.1 model on over 947 brain MRI reports in Dutch, focusing on variables like atrophy and microbleeds. The AI achieved near-perfect accuracy for categorical data-96% for medial temporal atrophy on the right and 87% for global cortical atrophy-while showing room for improvement with numerical data. The study highlights how few-shot prompting can enhance AI performance, boosting its ability to handle numbers by nearly 12 percentage points. This suggests that with the right strategies, AI could significantly reduce the time doctors spend on repetitive tasks like data extraction. However, challenges remain, particularly in accurately identifying specific lesion locations. Looking ahead, researchers will likely focus on refining these techniques to address remaining gaps. The potential for AI to automate data extraction from medical reports is enormous, offering a clearer picture of how these technologies can support healthcare professionals in the future.
OpenAI Prepares Biggest Update to ChatGPT
OpenAI is preparing a major update to ChatGPT. The update will transform the chatbot into a super app. It will combine coding tools and AI agents. The update is meant to push ChatGPT beyond conversation and into autonomous action. OpenAI's coding product Codex will get more resources. The company believes task-performing agents are more valuable than chatbots. The AI coding market is projected to reach $30 billion by 2031. OpenAI faces pressure to drive revenues higher and forge a path to profitability. The update will roll out in coming weeks. It will change ChatGPT's website and mobile apps to encourage users to try coding and other tools.
AI Agents Reject Traditional Sales Tools
AI agents prefer not to use traditional sales tools like Marketo, Outreach, and Salesloft. These tools are designed for humans to manage large volumes of emails and sales calls. AI agents can handle these tasks on their own and generate personalized emails and follow-ups in real time. By 2026, agents may handle 30% of these workflows, making traditional sales tools less necessary, and changing the way businesses operate will be next.