XNews

Anthropic Research & Infrastructure Findings

Anthropic Research & Infrastructure Findings

Anthropic released Frontier Red Team findings testing agents on blockchain smart contracts.

Key insights

Frontier Red Team simulated agent exploited smart contracts

Anthropic published new Frontier Red Team research detailing simulated tests where AI agents targeted blockchain smart contracts. The simulation identified a vulnerability worth $4.6 million. The research was conducted in collaboration with MATS Research and the Anthropic Fellows program and reported as part of Anthropic’s broader safety red‑teaming activities. (Post 15826)

Anthropic examined hardware FLOP sourcing shifts

A separate data visualization showed Anthropic’s compute FLOP sourcing composition changing between 2024 Q1 and 2025 Q3, illustrating shifts in GPU (orange) versus TPU (green) FLOP percentages. The graphic was presented as a comparison of Anthropic’s hardware reliance over the period. (Post 15680)

Research combined organizational partners for red‑teaming

Anthropic’s Frontier Red Team blog entry noted collaboration with external groups including MATS Research and Anthropic Fellows in conducting blockchain exploit simulations, indicating a multi-party research approach to red‑teaming. (Post 15826)

Source posts