Analysis
The October 9 Harness² preprint studies changes to agent prompts, tools and memory while model weights remain fixed. Its authors report improvements across three benchmarks and four configurations using GLM-5 and Gemini 3.5 Flash. The repository warns that executor network access is unrestricted, the Codex sandbox is disabled by default and public rubrics could be accessible. It describes research software, not an officially supported Google product.
The methodological question is what the reported improvement measures. In my assessment, a useful comparison would hold permissions and evaluation conditions constant while changing the harness. That would help distinguish better task organisation from advantages associated with the environment in which an agent operates.
Possible access to a rubric deserves examination, but a warning is not evidence that an agent accessed one or compromised an evaluation. A stronger account would document what information was available and what was actually used. Those details would make the reported gains easier to interpret without presuming misconduct.
For practical use, effectiveness and containment remain separate questions. An adaptive harness could become more capable without becoming easier to supervise. Comparisons under restricted permissions and tests on unfamiliar tasks would help assess that distinction. The supplied evidence leaves open whether the improvements persist under those conditions and does not establish safe self-modification.
Source: Google Research / arXiv reporting · 2026-10-09 · Primary source context · attributed account