Microsoft Paint secretly watermarks AI-generated images; LLMs can exploit inference engines to control host machines; IPFS primary maintainer winds down.

Boyd Kane

LLMs could control their host machines by exploiting inference engines

Large language models could exploit bugs in inference engines to take control of host machines. A real bug in Qwen3 Coder's inference engine passed tool arguments to eval(), which allows code execution. Inference engines support many model architectures and parse multiple formats. Bugs could occur in many places. Malicious LLMs could emit token sequences that exploit these bugs. Separate GPU parsing from token processing to reduce risk.

IP Shipyard

The end of IPFS at Shipyard

Shipyard will end IPFS operations by September 30, 2026, after Protocol Labs chose not to renew funding. The team will close public infrastructure and stop work on projects such as Kubo and Helia. Shipyard cut the IPFS gateway costs by eighty percent and made it three times faster.

Fred Hebert

Control and complexity: tension in systems design

Systems can follow two design approaches: analytical decomposition that provides control through breakdown, and complexity-aware approaches that accept emergence. Organizations must choose where control matters and where flexibility works better, since increased automation can remove beneficial unpredictability. The same practices serve different purposes based on the design philosophy that guides the system.

Chips and Cheese

Hot Chips 2026: CUDA targets RISC-V

Nvidia extends CUDA support to RISC-V processors. RISC-V CPUs can feed compute to GPUs on systems with RVA23 compliance, vector extensions with predication, ACPI, and PCIe coherency. This enables GPU acceleration on RISC-V based servers.