AI Coding News: August 18, 2026 — Claude Code Closes the Last NTLM Path Gap, Cursor Ships 3x-Faster Cloud Agent Builds
Anthropic shipped Claude Code v2.1.234 today, closing the last gap in last week's NTLM credential-leak fix and cutting the context cost of its built-in skill loading by roughly 8x. Cursor matched it with an infrastructure upgrade of its own, shipping Cloud Agent builds for 3x-faster agent startup. GitHub Copilot, OpenAI's Codex CLI, and OpenCode all stayed quiet since our last report. Here is what actually changed in AI coding tools this week, straight from the changelogs — with links to every source so you can verify it yourself.
Claude Code v2.1.234: Closing the Last NTLM Path Gap
The Claude Code changelog moved to v2.1.234 on August 17. The headline fix extends last week's Windows credential-leak patch: remote file reads, session restore, CLAUDE.md includes, workflow scripts, and file uploads now all reject Windows NT-namespace (\??\) paths too, closing the remaining pre-approval file-access routes that the NTLM leak vector could still reach after 2.1.233. The same release also delivers a major efficiency win — the context cost of loading the built-in claude-api skill dropped from roughly 200k tokens to about 25k by loading reference docs on demand, an approximately 8x reduction that leaves more context budget for the actual task.
- Security: remote file reads, session restore, CLAUDE.md includes, workflow scripts, and file uploads now reject Windows NT-namespace (\??\) paths, hardening the remaining pre-approval file accesses against the NTLM credential-leak vector.
- Reduced the context cost of loading the built-in claude-api skill from ~200k+ tokens to ~25k by loading reference docs on demand — roughly an 8x cut.
- Claude Code now continues your session automatically when a claude.ai usage limit resets; turn it off in /config ("Continue automatically at usage limit").
- New GitLab merge request badge in the footer and statusline: repos with a GitLab remote and an authenticated glab CLI show MR !N with draft/pending/green states.
- Fixed auto mode in very long sessions repeatedly re-checking and denying sandboxed commands' network access after the conversation had been compacted.
- Fixed session-scoped permission answers, including denies, being dropped when answering background subagent tool permission prompts.
- Claude is now told to use your account email only to identify you, not to send it to unrelated services unless you ask.
- Dozens of additional fixes landed across Remote Control, MCP diagnostics, markdown rendering, and the fullscreen TUI — see the full changelog for the complete list.
Cursor Ships Cloud Agent "Builds" for 3x-Faster Startup
Cursor's August 13 changelog entry introduces builds: pre-configured development environments that Cursor prepares in the background instead of setting up from scratch on every session. Agents now boot into an environment with the repository already cloned and dependencies already installed, which Cursor says gets agents to their first token roughly 3x faster. If a recent commit breaks the environment, agents fall back to the last successful build instead of stalling, and a new Builds tab in the Cloud Agents dashboard shows build status, logs, and which commit each agent used. Builds roll out automatically for new environments and are included at no extra cost with Cloud Agents.
Quiet Week for Copilot, Codex, and OpenCode
Not every tool ships something new every day. Checking each project's own changelog turned up nothing dated after our last report: GitHub Copilot's most recent entry is still Grok 4.6 landing on August 14, OpenAI's Codex CLI has not posted a new entry to its official changelog since reaching v0.147.0 on August 13, and OpenCode's latest tagged release is still v1.18.18 from the same day. Worth noting anyway: OpenAI's ChatGPT and Codex desktop app reached Linux in preview on August 11, with official .deb and .rpm packages for Ubuntu, Debian, and Fedora on x64 and ARM64, bundling Codex alongside ChatGPT and ChatGPT Work.
What This Means for Developers
Two threads run through this week's updates. First, security hardening is iterative, not a one-shot fix — Claude Code's NTLM leak took two releases to close across every pre-approval file-access path, a reminder to keep CLI tools on auto-update rather than treating last week's patch as the end of the story. Second, the competitive edge in agentic coding tools is shifting toward infrastructure efficiency rather than headline features: Claude Code's 8x cut to skill-loading context cost and Cursor's 3x-faster Cloud Agent builds both attack the same problem from different angles — making agents cheaper and faster to run at scale, not just more capable per token.
Resources & References
Comments
Share your thoughts and join the conversation