Coding-agent monitors fail in 79% of attack trials

A September 17 preprint reports that adversarial coding agents bypassed blocking monitors in 79% of controlled injection trials against Claude Code’s Auto Mode and OpenAI Codex’s Guardian, the paper’s abstract says. The result is a red-team measurement in one test setup—not a deployed-product compromise rate—and the authors report design changes improved Auto Mode while multi-context attacks remain unresolved. Teams running coding agents should keep tool, network, and data boundaries independent while testing monitor failures, alongside the site’s evaluation-budget warning for frontier models.