Coding-agent sandbox bugs let AI escape containment

Sandbox escapes in Claude Code, Codex and Cursor let coding agents break out of their containers and reach systems outside them, according to Accomplish founders Amit Avner, Or Hiltch and Guy Zipori, as reported by Upstarts Media.

The sandbox is what stands between an agent's mistake, or a prompt injection, and the actual machine it's running on. A leak there isn't cosmetic.

Cursor fixed its issue in about a week after a July report. OpenAI fixed two separate issues in August on a similar timeline and confirmed both on the record. Anthropic's fix took roughly 50 days and about 30 releases to land, and Anthropic declined to comment. Cursor also declined to comment on the record.

That's roughly 7x longer than the other two vendors, going by the report's own numbers. Neither company has said why. Thirty releases is a lot of surface for one patch to move through, and the piece doesn't say whether the gap was triage, complexity, or something else entirely.

Accomplish has a commercial interest in this finding, per the report. That doesn't make the bugs fake, but it's a reason to want the timeline confirmed independently rather than taken at face value.

Ask a vendor how fast they patch a sandbox escape once one's confirmed, not just whether one exists. On this report, the answers aren't the same.