Follow the latest coverage, related explainers and connected technology stories.
PrismML has launched the Bonsai 2 27B model, which compresses the Qwen3.8 27B model to 5.9 gigabytes while retaining 98% of its benchmark results, bringing the operation of advanced reasoning models closer to personal computers and potentially high-end phones. The company is betting on ternary weights to reduce memory requirements without significant performance loss.
Anthropic said that Claude models were used between December 2025 and August 2026 to develop software for conventional weapons, and in espionage, influence, surveillance, and fraud operations, in addition to unauthorized attempts to replicate artificial intelligence models. The company said it disabled accounts linked to these activities and shared information with public- and private-sector entities.
Anthropic said that five campaigns linked to Chinese artificial intelligence companies attempted to extract advanced capabilities from the Claude model through nearly 200 million interactions, including one campaign attributed to Alibaba and another to Moonshot AI. The company believes these methods could be used to produce training data for competing models, although the report presents the company’s allegations and the material provides no independent evidence supporting them.
CISA, NSA, and the FBI said that six Chinese artificial intelligence companies carried out extensive operations to extract outputs from advanced U.S. models through billions of tokens and millions of requests since late 2024. The agencies recommend improving monitoring of API abuse and sharing indicators of these campaigns among companies.
Perplexity has launched the Hybrid Compute feature for its Computer platform, allowing the intelligent agent to distribute tasks between advanced cloud models and local models running on Macs equipped with Apple silicon chips. The company says an on-device privacy gate scans sensitive data before any part of it is sent to the cloud, although the risks of misclassification and questions about usage data remain.