AI Daily — May 9, 2026
2026-05-09
openai.com
Running Codex safely at OpenAI
OpenAI published details on how Codex, their coding agent, is secured in production: sandboxed execution environments, network egress policies, human approval gates for sensitive operations, and agent-native telemetry for audit trails. The writeup targets enterprise adoption concerns around agentic coding workflows running autonomously in CI pipelines. It offers a concrete reference architecture for organizations evaluating coding agents in compliance-sensitive environments.
code.claude.com
Claude Code Week 19: Plugin loading from ZIPs and URLs, hard deny rules in auto mode
Claude Code's Week 19 update adds support for loading plugins from .zip archives and remote URLs via --plugin-dir and --plugin-url flags, enabling ephemeral plugin installs per session. Auto mode gains hard deny rules that unconditionally block specified actions regardless of any allow exceptions, strengthening guardrails for unattended operation. Hooks now also have visibility into the active effort level, allowing conditional logic based on the current task intensity.
code.claude.com
Claude Code Week 18: Native Windows support, ultrareview for CI, PR URL session resume
Claude Code Week 18 drops the Git Bash requirement on Windows, falling back to PowerShell as the default shell tool, broadening accessibility on Windows developer environments. The new `claude ultrareview` command brings cloud-based code review into CI pipelines and scripting contexts. Additionally, pasting a PR URL into /resume now automatically finds and restores the session that created that PR.
arxiv.org
ZAYA1-8B: Sub-1B active parameter MoE model competitive with larger reasoning models
Zyphra's ZAYA1-8B is a mixture-of-experts model with only 700M active parameters out of 8B total, trained entirely on AMD hardware using their MoE++ architecture. It claims to match or exceed DeepSeek-R1-0528 on several math and coding benchmarks despite its smaller active footprint, with reasoning data incorporated from pretraining onward via an answer-preserving trimming scheme. Post-training uses a four-stage RL cascade covering math/puzzle warmup, suggesting a practical path to high-efficiency reasoning models on non-NVIDIA stacks.