In the past hour, total liquidations across the network reached $15.31 million, with $2.00 million from long positions, with $13.31 million from short positions. BTC liquidation $7.37 million, ETH liquidation $2.51 million.

According to Cointelegraph, Immunefi CEO Mitchell Amador stated that AI models such as Claude Opus 4.8 and ChatGPT 5.5 contributed to a resurgence in crypto hacking activity in 2026. DefiLlama data shows that over $634 million was stolen from crypto platforms in April 2026, the highest monthly total since the $1.4 billion loss from the Bybit hack in February 2025. On April 19th, attackers transferred 116,500 RSETH, worth approximately $292 million, from the Kelp DAO's RSETH bridge. LayerZero stated that Kelp DAO's 1/1 decentralized validator network configuration created a single point of failure.

According to CoinWorld, Ryan Lee, Head of Developer Relations at MiniMax, announced that MiniMax Sparse Attention (MSA), a high-performance attention library for NVIDIA Blackwell (SM100) GPUs, has been officially open-sourced under the MIT license. MiniMax-M3 weights are expected to be released this Friday. MSA is used for reasoning with millions of tokens in context, filtering relevant key-value blocks and performing attention calculations only on selected blocks. Research shows that in a context of 1 million tokens, MSA reduces attention computation by 28.4 times compared to Dense GQA with the same configuration, and achieves 14.2 times pre-padding speedup and 7.6 times decoding speedup on an H800 GPU. The open-source version integrates C++ JIT and Cute-DSL implementations and supports multiple precision formats.

According to CoinWorld, AI startup Recursive has released experimental results for its research system. This system can automatically generate ideas, write code, run experiments, and validate them, exceeding publicly available best results in three benchmarks: fixed-budget training, nanogpt speedrun, and GPU kernel optimization. In a 5-minute nanogpt autoresearch training session, the system reduced the validation loss bitbb to 0.9109, shortening training time by approximately 23% (a 1.3x speedup). In nanogpt speedrun, the system reduced the training time to a specified loss from 79.7 seconds to 77.5 seconds. In the GPU kernel optimization benchmark SOL-execbench, the system improved the average SOL score on an NVIDIA B200 from 0.699 to 0.754, narrowing the gap to the physical limit by 18%. To prevent AI from exploiting vulnerabilities to inflate scores, the system incorporates multi-level correctness audits to filter out invalid speedups.

According to CoinWorld, Hermes Agent has launched a web-based configurator, allowing developers to create and configure agent roles in a single web interface. The configuration process has been simplified into five visual steps, covering agent identity naming and description, model provider and inference parameter settings, installation of built-in and optional skill libraries, configuration and testing of the model context protocol server, and a final check and preview. Developers can access the profiles page in the left navigation bar of the webpage and click the build button at the top to navigate to the /profiles/new path to begin building.