WORK WITH BUSINESS CLIENTS?

Sell AI employees under your own brand.

Coding Weekly AI News

November 9 - November 17, 2026

Weekly signal

Period: November 9–17, 2026. This week sharpened the operational plumbing for agentic coding tools (new runtime features, model-default churn, and local/CLI ergonomics) while a high‑visibility data‑exposure finding underscored product design risks when agent outputs become shareable web apps. Key developments matter for engineering teams, SREs and security.

What changed

  1. Anthropic updated Claude Code (v2.1.323, published Nov 12) with an experimental “Code Mode” where the agent emits JavaScript to call tools (opt‑in), tighter plugin version ranges, a 1M default context for previously‑unknown model IDs, and a large sweep of permission/sandbox fixes. This landed as an npm/GitHub release and changes how tool invocation and plugin compatibility operate in terminal coding agents.

  2. OpenAI shipped a Codex CLI release (rust‑v0.169.0, Nov 13) that adds multi‑tab side chats, opt‑in cloud/archive options, provider‑management from the terminal, and several sandbox/safety fixes relevant to local agent runs and CI integration. CLI ergonomics for agent sessions keep evolving rapidly.

  3. GitHub’s Copilot CLI released v1.0.110 (Nov 12), adding experimental masked responses for git/ssh/sudo prompts, accessibility and session UI improvements, and a compact operator that now respects session AI credit limits — important for teams running shared agent sessions or automations.

  4. Microsoft’s Conductor (agent/workflow orchestration) changelog (0.1.45, Nov 10) documents multi‑provider runner support (Copilot, OpenAI, Claude), per‑call secret delivery into runners, and changed default model selections as vendors retire older models — a signal that orchestration layers are standardizing execution isolation and model routing.

  5. Security: Claroc’s “GhostShares 2” research (reported by Ars Technica Nov 10) found tens of thousands of Grok‑built apps reachable via grok.me hostnames; many were published with “anyone with the link” and contained internal or sensitive data. Cloud Security Alliance noted this as part of a recurring AI app‑builder leak pattern — practical warning for teams that let agents publish apps or artifacts to the web.

What to do with it

  • Patch and pin: upgrade terminal/CLI agents (Claude Code, Codex, Copilot CLI) this week and pin model IDs in orchestration workflows to avoid implicit default shifts.
  • Harden publish flows: audit any agent feature that publishes web artifacts (Grok or similar). Search passive DNS (e.g., grok.me) for accidental exposure and set internal defaults to “Only you” or org‑invited. Test from a signed‑out browser.
  • Enforce runtime isolation: use Conductor/runner execution profiles or containerized runners and per‑call secret overlays so agent‑executed processes can't leak secrets into host environments.
  • Monitor limits & budgets: enable Copilot/Codex usage dashboards and make /compact and credit limits part of CI checks to avoid surprise consumption.
Extended Coverage

FOR AGENT BUILDERS

Run your agents in a persistent workspace.

Run Hermes and OpenClaw with saved context, connected tools and delegated coordination.