Anthropic details its engineering approach to containing the potential "blast radius" (potential for damage) of increasingly capable AI agents like Claude. The company identifies three primary risk categories: user misuse, model misbehavior (e.g., unexpected sandbox escapes), and external attackers (e.g., prompt injection).

To address these risks, defenses are applied to three main components: the environment in which the agent runs (sandboxes, VMs, egress controls), the model itself (system prompts, classifiers, training modifications), and the external content the agent can reach (tool permissions, input inspection). Human-in-the-loop supervision is found to be fallible due to "approval fatigue."

Anthropic employs three distinct isolation patterns tailored for different products: ephemeral gVisor containers for claude.ai code execution, a human-in-the-loop sandbox on the user's machine for Claude Code (with OS-level sandboxes to mitigate approval fatigue), and a full virtual machine (VM) for Claude Cowork, designed for non-technical users to provide absolute and always-on boundaries.

The company also highlights missed risks, including code execution before user consent, credential exfiltration via user-driven prompt injection, data exfiltration through approved domains, and VM isolation hindering endpoint detection software visibility. Key principles emphasized are designing for containment at the environment layer first, matching isolation strength to the user's capacity for oversight, and being wary of custom components, trusting battle-tested primitives over bespoke solutions.

👉 Read the full article on Anthropic

A startup named Human Archive has secured $8.2 million in funding from prominent investors including Wing Venture Capital, NVP Capital, Y Combinator, and angel investors from OpenAI, Nvidia, Google, and Meta. The company is engaged in collecting AI datasets by attaching head-mounted and wrist-mounted cameras to workers in diverse environments globally, including those on gig economy platforms in India.

Co-founders Rushil Agarwal and Raj Patel aim to build "foundational datasets" for modeling human sensimotor intelligence at scale, creating 3D datasets from head cameras and 2D hand movement datasets from wrist cameras. They believe their technology will become "foundational infrastructure for automating manual labor, increasing global abundance, and advancing our understanding of human intelligence itself."

The article raises concerns that this data collection is intended to train AI models to automate manual labor and potentially replace human workers with robots. It draws parallels to a viral video in India purportedly showing garment workers wearing head-mounted cameras in factories.

👉 Read the full article on Gizmodo

The content of the provided URL was inaccessible. However, the article title, "The Electric Ferrari Luce Is Finally Here," suggests that Ferrari has made a significant entry into the electric vehicle market with the unveiling of its first fully electric model, named "Luce."

This announcement would mark a pivotal moment for Ferrari, a brand historically synonymous with high-performance internal combustion engines, as it embraces electrification. The Luce is anticipated to combine the brand's iconic design, unparalleled performance, and cutting-edge technology with an electric powertrain, setting new benchmarks in the luxury EV segment.

The arrival of this EV reflects the broader automotive industry's shift towards electrification and Ferrari's commitment to adapting to a sustainable future while preserving its heritage. Specific details regarding its specifications or market availability remain unknown due to the inaccessibility of the original article, but its mere existence would generate considerable attention.

👉 Read the full article on Wired