OpenAI Describes the Monitors That Watch Its Own Agents, and Where They Stop
An October 29 OpenAI post sets out a four-layer system that checks every reinforcement learning run and evaluation for models at least as capable as GPT-6 Astra. Its most useful details are the gaps: paging is off for internal Codex traffic, and the external monitor cannot pause every API conversation.
Nemotron 3.5 Transcribe: 4.3 Billion Active Parameters, 34 Billion to Hold in Memory
NVIDIA released an open-weight speech model on October 28 that transcribes, labels speakers, translates and answers questions about recordings up to an hour long. The model card's hardware table shows why active parameters do not decide which GPU can run it.
Oracle Will Rent Out Google TPUs Next to NVIDIA Racks
At Oracle AI World on October 27, Oracle said Google TPUs will be available on OCI in 2027, the first time Google's accelerators will be offered by another hyperscale cloud. The same day, NVIDIA Vera Rubin NVL72 systems and Vera CPUs became generally available on OCI.
Framework Watch: An Agent Exposed Over MCP Sent Its Hidden Reasoning to Clients
ADK Python stopped to_mcp_server from returning a model's thought parts to MCP clients. Claude Code 2.1.314 adds a setting that blocks writes outside the working directories in every permission mode. Codex 0.166 keeps approval settings across resume and fork, and PyRIT fixed a Crescendo refusal check that was judging the wrong text.
Research Radar: A Self-Improving Agent Can Delete a Safeguard Without Its Acceptance Gate Noticing
Papers submitted to arXiv on October 28 find that acceptance gates let overlapping safeguards erode unseen, that injected runtime feedback steered commercial coding agents in about 60 percent of trials, and that malicious CLAUDE.md files had no consistent effect on code security. Two more look at agent-written papers and agent repairs.
Trend Watch: The Harness Is Becoming a Product of Its Own
Microsoft now offers the GitHub-built Copilot harness as a managed runtime in Foundry. Two preprints the same week find that the harness, not only the model, decides results: lab harnesses scored zero on a long audit task away from their home model, and a built-in review policy changed which vulnerabilities a model reported.