AiPulse — Issue No. 91

Meta announced that it has removed a clause added to its commercial terms of use that gave the impression that the content of sites with Meta business accounts could be used to train artificial intelligence models. The company stated that this change does not alter how web browsers are used and is limited solely to advertising products.

Kamu Siber Güvenlik Derneği, Ankara Yıldırım Beyazıt Üniversitesi ev sahipliğinde yapay zeka ajanlarının siber güvenlikteki kullanımını ve risklerini inceledi. Katılımcılar, istem enjeksiyonu saldırılarının etkilerini ve güvenlik önlemlerini uygulamalı olarak denedi.

Researchers from Google Quantum AI, DeepMind, and MIT have developed a modular architecture that reduces the physical qubit requirement by a factor of 30 compared to standard codes. This approach enables achieving high coding rates without the need for complex intra-chip connections.

ARPA-E allocated a budget of 37 million dollars to one research team for three years. The program aims to achieve a speed 100 times that of classical computing for energy applications.

Researcher Karen Antorveza Paez combined the biocementation process, which converts sand into a solid structure without heating, with digital manufacturing. This method offers a more sustainable alternative to traditional cement production, which consumes high amounts of energy.

Muse Spark Open, released by Meta, has risen to become the most powerful open-weight model of US origin. This move halved the performance gap with Chinese rivals, altering the competitive balance among Western labs.

Tencent Hunyuan released the Hy Memory 2.0 agent memory framework, the Hy-Memory-30A3B model, and the MemHub application as open source. The company states that the Hy-Memory-30A3B model outperformed Claude Opus 5 in benchmarks.

Palo Alto-based Elorian has introduced Glimpse, a model that reasons by re-examining images. The model is offered at a significantly lower price than OpenAI's GPT-6 Astra and outperforms its competitor in some comparisons.

A team from Harvard and Anthropic demonstrated that Claude improved resource efficiency in 7 out of 8 tasks by redesigning quantum circuits. However, the notable 17-fold improvement was measured against a previous design using fewer qubits, not against the best design across all metrics.

How Developers Secure AI-Generated Code: 5 Security Best Practices

Developers are applying five core security principles using a 'shift-left' approach to prevent security vulnerabilities in AI-generated code.