Report: GLM-5.3 reportedly finds a serious flaw in Cursor
Z.ai's open cyber model hits 84.5% on CyberGym and is probing the IDEs you code in — plus Grok 4.6 spreads across tools and an open post-transformer.
Copy markdown
The flaw is real news, but unconfirmed
Security researcher Joshua Saxe reports GLM-5.3 turned up a “serious vulnerability” in Cursor while probing it. No technical details, CVE, or Anysphere patch have surfaced yet — treat it as a signal, not an advisory.
The cyber scores behind it
GLM-5.3 posts 84.5% on CyberGym and 54.4% on ExploitBench, up from 24.4% for GLM-5.2. Z.ai says its models have flagged 2,436 vulnerabilities across 269 projects, 1,097 rated critical or high. Automated bug-hunting at open-frontier scale is now cheap.
Your AI IDE is the attack surface
The documented risk isn't the model — it's zero-click prompt injection. The DuneSlide flaws (CVE-2026-50548/50549, CVSS 9.8) let a poisoned MCP server or web result escape Cursor's sandbox to run code; they're patched in Cursor 3.0+. Stay updated and treat anything your agent reads as untrusted.
Elsewhere: Grok 4.6 lands in your editor
xAI's frontier coder ($2 / $6 per million tokens, 500K context, about half the price of rival frontier models) is now selectable in Cursor with a first-week 2x-usage bonus and live in GitHub Copilot, plus OpenRouter and Vercel.
Elsewhere: an open post-transformer to poke at
Pathway open-sourced “Dragon Hatchling” (BDH); its 150M BDH-CQ hits 29.5% on ARC-AGI-1 at $0.0007 per task — roughly 11x cheaper per task than GPT-5.6 Luna (which scores 34.2%) — by reasoning in a recurrent latent state instead of a chain-of-thought scratchpad. The code is on GitHub.