Welcome to our monthly newsletter, which we have prepared to bring you current developments and informative content on Artificial Intelligence Security!
🙋🏻♀️Editor's Note:
Our website aisafetyturkiye.org is under construction. During this period, you can reach us via aisafetyturkiye@gmail.com.
ASEAI '25 brings together leading AI Security and ethics researchers. Don't forget to submit your article by the extended deadline for a chance to attend this important conference ahead of the AI Action Summit in Paris!
Organized by the Oxford AI Safety Initiative, the 2-week intensive bootcamp aims to rapidly develop ML security skills. This training will focus on topics such as creating the GPT-2 Small model, learning interpretability techniques, understanding RLHF, and reproducing important research articles.
A 3-6 month full scholarship research program to advance the safe and beneficial development of Artificial Intelligence with the world's leading AI Security organizations – such as CHAI, FAR.AI, Mila Quebec and Conjecture.
BlueDot Impact organizes 1-month crash courses on Transformative AI and Introduction to AI Adaptation. Don't forget to catch the next application period!
Participants in this 3-day spring school will examine the social, economic, technical and legal dimensions of Generative AI through a dynamic mix of lectures, panel discussions and hands-on activities.
OpenAI's newly released system board (security evaluation) of the o1 model highlights the review conducted with external experts Apollo Research and highlights key concerns. Tests showed examples of 'planning', where o1 secretly worked towards its own goals rather than following user instructions. While this type of behavior is also present in other advanced AI models, o1 was seen more frequently, exhibiting more sophisticated patterns in pursuit of alternative goals.
New AI models and performance metrics are released almost daily, and it becomes difficult to keep track of how to compare them to each other or which comparison you should trust. Epoch AI aims to create a central site to keep key AI performance benchmarks together.
OpenAI has launched its long-awaited new artificial intelligence model o1 as a full version.
The model, which was previously in the testing phase under the code names o1-preview and “Strawberry”, brings significant improvements such as improved performance in coding and mathematical tasks, image-based reasoning capabilities and shorter and more concise answers.
OpenAI also announced that it is working on future updates such as web browsing integration, advanced file upload features and new functions.
Anthropic and Palantir Technologies have joined forces with Amazon Web Services (AWS) to announce a strategic partnership that will bring Claude AI models to secret US intelligence and defense agencies.
Claude models, which will use AWS infrastructure on Palantir's Artificial Intelligence Platform (AIP), will have Impact Level 6 (IL6) security accreditation.
This collaboration will deliver critical capabilities such as rapid processing of complex data sets, advanced pattern recognition, improving document review processes, strengthening decision support mechanisms, and integration with existing analysis workflows.
Amazon invested an additional $4 billion in Artificial Intelligence company Anthropic, increasing its total investment to $8 billion.
With this new investment, AWS has been identified as Anthropic's main training partner. Anthropic has committed to using Amazon's custom Trainium and Inferentia chips in future model development.
The integration of Claude AI into Amazon's Alexa ecosystem is planned for 2025.
Researchers from the University of Pennsylvania have identified significant security vulnerabilities in robot systems that work with Large Language Models (LLMs).
These vulnerabilities indicate that systems can be manipulated to perform dangerous or unauthorized actions. The study involved a simulated autonomous vehicle violating traffic rules, a wheeled robot (Jackal) helping to plant potential explosives, and a four-legged robot (Go2) being used for unauthorized surveillance and access to restricted areas.
Researchers uncovered these vulnerabilities using the RoboPAIR system and various LLMs such as Nvidia's Dolphin model and OpenAI's GPT-4 model.
The Biden-Harris Administration announced the implementation of more than 100 measures from the landmark Executive Order on Artificial Intelligence.
This comprehensive progress report details key achievements in areas ranging from safety and security to workforce development and international leadership.
Advances in AI system monitoring, testing and evaluation, biosecurity measures, worker and consumer protection, innovation and research, government enforcement, and international leadership reinforce the United States' global leadership in responsible AI development.
The Biden-Harris Administration announced that it has implemented more than 100 measures within the scope of the Executive Order on Artificial Intelligence and published an annual progress report.
According to the report, significant improvements have been made in many areas, from safety and security to workforce development and international cooperation.
Audit of Artificial Intelligence systems, testing and evaluation processes, biosecurity measures, protection of employee and consumer rights, innovation and research studies, Artificial Intelligence applications in government institutions and global collaborations are among the focal points of the report.
JOB POSTINGS 👩🏻💻
To explore new opportunities in the field of Artificial Intelligence Security
You can take a look at 80.000 Hours' job postings page!
