WORK WITH BUSINESS CLIENTS?

Sell AI employees under your own brand.

Coding Weekly AI News

November 9 - November 17, 2026

Weekly signal

Period covered: November 9–17, 2026. This week produced concrete operator‑level changes in agentic coding stacks (terminal agents, CLIs, and orchestration layers) plus a high‑impact security disclosure about AI‑built web apps. The practical lesson is simple: agent features are moving from “assistant” to “actor” — they execute, publish, and change systems — so builders must treat them like another service in the deployment and security lifecycle.

What changed

Claude Code (Anthropic) — release v2.1.323 (Nov 12).
Anthropic pushed a sizeable terminal/agent release that adds an experimental Code Mode where Claude generates JavaScript to call tools (opt‑in), introduces engines version ranges for plugins (plugin.json compatibility), expands the default context size for unknown model IDs to 1M, and applies a broad set of permission and sandbox fixes. The release notes show many small but safety‑adjacent fixes (Bash sandbox prompts, subagent behavior, background session sign‑in handling) that materially affect how agent tasks run against a repo in CI or on developer machines. For teams embedding Claude Code into pipelines, Code Mode changes the attack surface for tool calls and plugin behaviors and the new plugin version constraints change compatibility guarantees.

OpenAI Codex CLI — rust‑v0.169.0 (Nov 13).
Codex’s terminal agent continues to bake in collaboration and cloud session features: multiple side‑chat tabs, opt‑in cloud connection/archiving, provider onboarding from the terminal, and important sandbox and Windows fixes. If you use Codex in developer tooling or CI, the new provider management and cloud archive options change where sessions and artifacts live (local vs gated cloud).

GitHub Copilot CLI — v1.0.110 (Nov 12).
Copilot CLI’s release adds experimental masked password handling for agent‑driven git/ssh/sudo prompts, session accessibility improvements, and layout/compact fixes that respect AI credit limits. This is practical: teams using Copilot automations that touch remotes or need credentials can test the new masked dialog, and the compact/credit behavior helps avoid accidental overrun of pooled budgets.

Microsoft Conductor — changelog 0.1.45 (Nov 10).
Conductor’s recent notes show orchestration hardening: execution realms/profiles (Docker or Azure Container Apps) for isolated agent runs, remote runners supporting /interrupt, multi‑provider execution with per‑call agent‑scope secret delivery, and explicit warnings about implicit model defaults changing as providers retire models. That changelog signals a maturation: orchestration layers are expecting multi‑provider agents, need to deliver secrets per run, and must make model choices explicit to avoid silent budget or runtime surprises.

Security disclosure: Grok Build / GhostShares 2 (Claroc, reported Nov 10; coverage Nov 10–11).
Security researchers enumerated grok.me hostnames and reported tens of thousands of Grok‑built apps reachable without sign‑in; a non‑trivial fraction contained internal or sensitive data, including some built by the vendor’s own staff. Claroc’s findings (summarized by Ars Technica and flagged by CSA) show a recurring class of product risk: publish/share flows that put secrets in places discoverable through passive DNS and thereby expose internal data. This is directly relevant to coding agents that scaffold, deploy, or publish web artifacts produced by agent runs.

Implications and analysis

  1. Operational risk from implicit model and runtime defaults.
    Conductor’s changelog and several vendor releases show providers are retiring older models and shifting defaults; orchestration layers are already changing default model IDs. If you rely on implicit model defaults in agent workflows, you risk silent cost and behavior changes (model retirement, different reasoning effort, or different context pricing). Pin models explicitly in pipelines and surface retirements to your deploy and change control process.

  2. Agent tooling is consolidating complex behaviors into CLIs and terminal UIs.
    Codex and Claude Code continue to add cloud/offload options, provider configuration, and plugin ecosystems. That convenience also centralizes risk: a single background session or misconfigured gateway can cause unattended runs or data exfiltration. Treat agent sessions like running services: monitor, log, rotate credentials, and audit their permitted MCP/tool set.

  3. Publishing generated apps is a distinct class of supply‑chain risk.
    The Grok findings show that UX choices around where the randomness lives in a URL and the default access selection can convert an internal tool into an enumeratable public service. When agents produce web apps or artifacts that get hosting addresses, security teams must own discovery (passive DNS sweeps), auditing, and default access settings.

  4. Secrets & least privilege at execution time.
    Conductor’s env_overlay/per‑call secret delivery and remote runner isolation are the right moves: secrets should be injected only into the agent process that needs them and should not persist in a runner’s environment or artifacts. Make per‑call secrets, ephemeral tokens, and runner sandboxing standard operating procedure.

What to do with it (practical next steps)

Immediate (this week)

  • Upgrade/pin agent tooling: update Claude Code to v2.1.323 (or the latest patch), Codex CLI to rust‑v0.169.0, and Copilot CLI to v1.0.110 (or newer) in developer images and CI, test in a staging environment, and run smoke tests that validate sandboxes and permission prompts. Pin explicit model IDs in orchestration manifests to avoid implicit default changes.
  • Audit published artifacts: run a passive‑DNS sweep (or ask security team to do so) for any builder hostnames your org’s developers might use (for example grok.me, or other vendor share domains). Manually open published apps from an incognito window to confirm their access setting; unpublish or change to org‑invited/Only‑you if internal. Treat any agent‑publish flow as a pre‑release checklist item.

Short term (1–4 weeks)

  • Enforce per‑run secret delivery and containerized runners: require orchestration layers (Conductor, workflows) to use execution profiles or Docker runners and deliver secrets via per‑call overlays, not environment persistence. Add CI checks that fail when an agent run attempts to access secret stores without a validated policy.
  • Instrument budgets and /compact behaviour: wire Copilot/Codex usage telemetry into your cost dashboards and alert if sessions approach pooled budget thresholds. Make /compact and other token‑management commands part of nightly maintenance.

Policy & governance

  • Update developer guidance for agent publishing: add a checklist that covers sharing defaults, passive‑DNS discoverability, and “don’t store passwords in app code” rules. Require security approval for any agent flow that publishes web endpoints or automates sign‑in logic.
  • Review plugin marketplaces and plugin.json engine ranges: if you run private plugin registries or gateways, require plugin.json compatibility ranges and validation to be part of CI checks so an incompatible plugin isn’t auto‑installed into a running agent.

Sources Anthropic — Claude Code release v2.1.323 (GitHub release notes). https://github.com/anthropics/claude-code/releases/tag/v2.1.323 OpenAI — Codex CLI release rust‑v0.169.0 (GitHub release notes). https://github.com/openai/codex/releases/tag/rust-v0.169.0 GitHub — Copilot CLI release v1.0.110 (GitHub release notes). https://github.com/github/copilot-cli/releases/tag/v1.0.110 Microsoft — Conductor changelog (0.1.45, Nov 10, 2026). https://raw.githubusercontent.com/microsoft/conductor/main/CHANGELOG.md Dan Goodin, Ars Technica — “Grok Build coding agent leaks oodles of confidential info” (Nov 10, 2026). https://arstechnica.com/security/2026/11/grok-build-coding-agent-leaks-oodles-of-confidential-info-some-belonging-to-spacexsi/ Cloud Security Alliance — Research note: "Public by Default: Exposed Grok Apps and the AI App‑Builder Leak Pattern" (Nov 11, 2026). https://labs.cloudsecurityalliance.org/research/csa-research-note-grok-build-apps-exposed-by-default-2026111/

If you want, I can: (A) produce a short checklist you can drop into an internal runbook to audit agent publish flows and passive DNS; (B) generate a PR template that forces model pinning and execution profile selection in your repo's workflow manifests; or (C) run a prioritized upgrade plan for the specific agent versions you use.

Weekly Highlights

FOR AGENT BUILDERS

Run your agents in a persistent workspace.

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