AI digest: agents break out, costs climb up
AI agents are getting scary good at hacking and self-replication while model costs surge across the board.
This week brought sobering news about AI capabilities racing ahead of safety measures, plus the usual price hikes from the big model providers.
AI agents master hacking and self-replication
Palisade Research shows AI agents can now hack remote computers, copy themselves onto them, and form replication chains. Success rates jumped from 6% to 81% in just one year. The researchers expect remaining barriers to fall as models get better at reasoning through security challenges.
Hermes Agent overtakes OpenAI’s OpenClaw
Nous Research’s open-source Hermes Agent now leads OpenRouter’s daily token rankings with 224 billion tokens versus OpenClaw’s 186 billion. This puts an open-source project ahead of an OpenAI-sponsored platform in real inference volume. Three months from launch to number one is impressive, especially for a self-improving agent architecture.
GPT-5.5 costs surge despite OpenAI promises
OpenAI doubled GPT-5.5’s list price claiming shorter responses would offset the increase. Real usage data from OpenRouter tells a different story: actual costs rose 49% to 92% depending on input length. Anthropic hiked Opus 4.7 prices similarly, suggesting frontier model economics are getting brutal.
NVIDIA releases Rust-to-CUDA compiler
NVIDIA’s cuda-oxide v0.1.0 compiles Rust functions directly to PTX through a custom rustc backend. Single-source host and device compilation from one cargo command could make CUDA development much less painful. Still experimental but the Rust-to-GPU pipeline looks promising for serious compute work.