AI Coding News: August 15, 2026 — Claude Code Patches a Windows Credential Leak, Copilot Adds Grok 4.6
Anthropic closed a Windows credential-leak vector in Claude Code today, GitHub Copilot added two more frontier models to its lineup, OpenAI's Codex CLI landed its promised Agent Plugins support, and OpenCode quietly fixed a pair of provider-routing bugs. Here is what actually changed in AI coding tools on August 15, 2026, straight from the changelogs — with links to every source so you can verify it yourself.
Claude Code 2.1.233: A Windows Credential-Leak Fix and a New GitLab Integration
The Claude Code changelog moved to v2.1.233 on August 14, and the headline item is a security fix: Windows paths using the NT \??\ device prefix could bypass UNC path validation, a vector that could leak NTLM credentials. Anyone running Claude Code on Windows should update immediately. The release also adds GitLab merge request URLs to the --worktree flag and the claude agents view (MRs now display as !N, matching how GitHub PRs already show as #N), plus an opt-in memory cgroup limit for Bash tool commands on Linux via CLAUDE_CODE_TOOL_MEMORY_LIMIT.
- Security fix: Windows NT \??\ device-prefix paths could bypass UNC validation, an NTLM credential-leak vector — patched in 2.1.233.
- GitLab merge request URLs now work with --worktree and the claude agents view, shown as !N.
- New CLAUDE_CODE_TOOL_MEMORY_LIMIT env var caps Bash tool memory via Linux cgroups; CLAUDE_CODE_WEBFETCH_CACHE_TTL_MS configures the WebFetch session cache.
- Reliability fixes: cloud sessions no longer get marked "lost" during shutdown while a permission prompt is pending, MCP v2 stops endlessly reopening subscriptions on serverless hosts, notification hooks fire again for prompts in Desktop and VS Code, and idle Linux sessions with sandboxing no longer peg the CPU at 100%.
- Notable behavior change: TaskCreate/Get/Update/List and TodoWrite are now disabled by default on Opus 4.8, Sonnet 5, and newer models. Set CLAUDE_CODE_ENABLE_TODO_TOOLS=1 to restore them if your workflow depends on them.
- 2.1.232's Bash permission changes for Cygwin-style symlinks and input redirections were reverted after causing regressions.
GitHub Copilot Adds Grok 4.6 and Gemini 3.7 Flash
GitHub expanded Copilot's model lineup twice in two days: Gemini 3.7 Flash landed on August 13, followed by Grok 4.6 on August 14, giving developers two more options alongside the existing Claude and GPT models across VS Code, the CLI, and the Copilot app. This follows last week's rollout of Agent Plugins 1.0 support in VS Code, Copilot CLI, and the Copilot app, and a reminder that several older models are being deprecated across all Copilot surfaces on September 1. See the full Copilot changelog for weekly VS Code, CLI, and desktop releases.
OpenAI Codex CLI v0.147.0: Agent Plugins and Safer Secret Handling
OpenAI's Codex CLI reached v0.147.0 on August 13, and it is the release where the cross-vendor Agent Plugins standard shows up in practice: Codex can now install portable Agent Plugins and search across local, personal, workspace, and remote plugin catalogs. The same build adds a new --approve-for-me flag for automatically reviewed approvals, lets you import Cursor-managed skills and sync imported Claude and Cursor conversations without creating duplicates, and ships opt-in support for the MCP 2026-07-28 protocol with paginated discovery and non-blocking server startup. On the security side, Codex now redacts secrets and complete bearer tokens from displayed commands and replayed conversation history. The prior patch, v0.146.1 on August 5, applied safer automatic-review defaults for cyber-capable models. Full history is in the Codex changelog.
OpenCode v1.18.18: Provider Fixes for Kimi and xAI
OpenCode, the open-source, provider-agnostic coding agent from SST, shipped v1.18.18 on August 13, correcting system prompt selection for the official Moonshot and Kimi providers and fixing xhigh reasoning-effort configuration for xAI models. The prior release, v1.18.17 on August 12, improved session compaction to keep complete recent turns with clearer summaries for smaller models, added reasoning variants for MERGE Gateway models, enabled PDF attachments for GitHub Copilot models that advertise the capability, and corrected routing for the Muse family of models. See the full release notes for details.
What This Means for Developers
Two threads run through today's updates. First, security hardening keeps shipping quietly alongside features — the Windows NTLM leak in Claude Code is a reminder to keep CLI tools on auto-update, since these are exactly the fixes you don't want to be late to. Second, model choice and plugin portability are converging: Copilot's Grok and Gemini additions and Codex CLI's new Agent Plugins support both point toward a market where the agent and the model are decoupled from the underlying skill or plugin ecosystem. If you rely on Claude Code's TodoWrite tool in scripted workflows, check whether you're on Opus 4.8, Sonnet 5, or newer — that tool is now off by default and needs an explicit environment variable to re-enable.
Resources & References
Comments
Share your thoughts and join the conversation