An AI model found a serious security flaw in Cursor — what it means for teams betting on AI coding tools

Z.ai's newly released GLM-5.3 model flagged a potentially serious vulnerability in the Cursor code editor within days of launch, part of a broader pattern of AI models now finding security flaws in the very AI coding tools founders are commissioning apps with.

Z.ai released its GLM-5.3 model on 14 August 2026, and its headline result arrived almost immediately: security researcher Joshua Saxe flagged that the model had identified a potentially serious vulnerability in Cursor, one of the most widely used AI code editors, involving Cursor’s Rust/Electron architecture and a risk of arbitrary file write. Cursor’s team is working on a fix; the flaw was disclosed privately rather than published. What makes this newsworthy isn’t the individual bug — IDEs get vulnerabilities disclosed all the time — it’s the benchmark context. GLM-5.3 scored 84.5% on CyberGym, a vulnerability-discovery benchmark, ahead of comparable frontier models, and its score on ExploitBench (which measures a model’s ability to reason through and actually execute an exploit) more than doubled from its predecessor, to 54.4%. Z.ai says the improvement came from post-training reinforcement learning specifically on security tasks — and that its models have now collectively surfaced 2,436 vulnerabilities across 269 projects since the prior release.

The pattern to notice: AI models capable of writing production code are increasingly also capable of finding — and, in principle, exploiting — flaws in the tools other AI models run inside. Cursor, Claude Code, GitHub Copilot and similar tools now sit deep inside real company infrastructure, often with repository access, shell access and the ability to run commands unattended. A vulnerability in the editor itself, not just in the code it generates, is a different category of risk than “the AI wrote a bug” — it’s “the tool you gave repo and terminal access to has an exploitable hole.”

So what

This isn’t a reason to avoid AI coding tools — it’s a reason to treat which one you standardise on, and how it’s configured, as a real engineering decision rather than a personal preference. Teams commissioning software built with AI-assisted workflows should ask their delivery partner what tooling they run, what access it has, and how quickly it patches. We build with these tools deliberately, with sensible limits on what any agent can touch in a client’s environment — see our AI-assisted development approach, or get in touch if you want a straight answer on how a project would actually be built.

← All research

More research.

GitHub Copilot adds Supabase, Webflow, Wix and five more connectors in public preview

GitHub Copilot now connects to eight more services, including Supabase, Webflow, Wix and Mixpanel, through verified MCP connectors in public preview. They join more than 25 existing ones.

Read ›

Gemini 4 Neon is generally available in GitHub Copilot, aimed at lower-cost agentic coding

Gemini 4 Neon is now generally available in GitHub Copilot for agentic coding. Google says its quality is close to Gemini 4 Carbon at a lower token cost per task.

Read ›

Claude Code 2.1.328 brings auto mode to Bedrock, Vertex and Foundry, and narrows /code-review to bugs

Claude Code 2.1.328 extends auto mode to Claude Opus 4.6 and Sonnet 4.6 on Amazon Bedrock, Google Cloud and Microsoft Foundry. /code-review now reports only bugs and CLAUDE.md violations.

Read ›

GitHub Copilot Evals: repeatable tests for AI agents, now in public preview

GitHub Copilot Evals lets teams write repeatable tests for AI coding skills, custom agents and MCP tools. It checks both the result and the path the agent took, and is in public preview.

Read ›

GitHub Copilot can now drive your browser: browser use enters public preview

GitHub Copilot browser use lets an AI coding agent control Chrome to debug web apps and run browser tests. It is in public preview across the CLI, the Copilot app, VS Code and JetBrains.

Read ›

Claude Code 2.1.327 refuses some symbolic links and lists broken setup in /status

Claude Code 2.1.327 tightens file and package permissions again and adds a /status list of broken agent, skill and hook settings. A short read for teams using AI coding tools on a live product.

Read ›

Start a project

Building with AI? Let's talk.

Tell us what you are building. We'll give you an honest view of the right approach.

Start a projectSee our work