NVIDIA announced the Open Agent Safety Platform, combining OpenShell—an Apache-2.0 sandboxed agent runtime that constrains files, networks, tools, processes and credentials—with Sentry, a hardware-isolated watchdog/reference design for BlueField-4 DPUs. It is a notable shift toward out-of-band agent enforcement and monitoring, but the fuller Sentry design is optimized for NVIDIA Vera/BlueField systems and the announcement provides no independent effectiveness evaluation. Sources: https://developer.nvidia.com/blog/nvidia-open-agent-safety-platform-a-reference-for-continuous-in-silicon-agent-monitoring/ ; https://github.com/NVIDIA/OpenShell
GitHub made Anthropic’s Claude Sonnet 5.5 generally available across Copilot plans and coding surfaces, including VS Code, Copilot CLI/cloud coding agent, JetBrains, Xcode and github.com (gradual rollout; usage-based at provider list pricing). GitHub says its early testing matched Sonnet 5 on coding tasks with fewer steps, tokens and tool calls and faster completion; this efficiency comparison is vendor-reported, with no independent benchmark linked. Source: https://github.blog/changelog/2026-09-28-claude-sonnet-5-5-in-github-copilot
A new practitioner report describes a genomics MCP workflow in which a Claude agent read 182 values from 26 public ENCODE files; a model-free replay matched the extracted values and review corrected parts of the agent’s interpretation. This is a useful, concrete example of validating tool-mediated research against replayable data, but remains a single practitioner report rather than an independent evaluation. Sources: https://www.rewire.it/blog/genomics-mcp/ ; https://www.reddit.com/r/AI_Agents/comments/1woa0r2/weekly_thread_project_display/