Three of the most widely used AI coding tools quietly had leaky sandboxes this summer, and how fast each vendor fixed them looked nothing alike. A Tel Aviv-based stealth security startup called Accomplish, founded by Or Hiltch, Amit Avner, and Guy Zipori, privately flagged sandbox vulnerabilities to Anthropic, Cursor, and OpenAI earlier this year, then went public with the findings this week. Accomplish’s CTO, Hiltch, put it plainly: “There’s a lot of talk about security now.”

A sandbox is the isolated environment meant to stop a coding agent from touching the rest of a developer’s machine or a company’s data while it works. Accomplish’s own writeup names its Claude Code finding “Beltdown,” and it lands in a pattern security researchers have documented repeatedly in 2026, including a “Configuration-Based Sandbox Escape” class flagged across Claude Code, Codex, and Gemini CLI in the spring, and a related set of vulnerabilities across Cursor, Codex, Gemini CLI, and Antigravity reported over the summer.

What set this disclosure apart was the response gap. A bug Accomplish flagged to Cursor in July, and two flagged to OpenAI, were each fixed in about a week. A similar issue flagged to Anthropic roughly two months earlier took about 50 days and around 30 separate software releases before a patch shipped. Anthropic and Cursor did not respond on the record to requests for comment; OpenAI said it addressed both of its flagged issues in August and is continuing to tighten controls on where its agents can write files.