Pattern: Novee Says the Vulnerable Layer Is the Harness, Not the Model

Novee Security·medium signalXBlueskyLinkedInCopy link

Novee researcher Elad Meged frames the Black Hat findings as a repeatable pattern rather than three unrelated bugs: the flaws live in the harness — the surrounding code managing tool permissions, execution, and sandboxing — not in the models. The recurring shape is one component marking attacker-influenced state as safe and another consuming it with greater authority. Notably, the flaws were found using default configurations on public repositories, meaning exposure was live rather than theoretical; this argues for threat-modeling your own agent scaffolding (hooks, wrappers, worktree isolation, CI runners) with the same rigor you'd apply to the model's outputs.