Wiz: GhostApproval breaks AI coding sandboxes

Send us Fan Mail AI coding assistants are introducing a new class of risk by reusing old attack techniques in ways most teams are not prepared for. A recent finding shows these tools can be tricked into writing outside their sandbox, while still showing safe paths to users, breaking the trust model many teams rely on. This matters because AI agents are now embedded in developer workflows with real access to code and systems. If their guardrails are unreliable, they become a new execution layer that must be treated as untrusted. At the same time, regulators in Europe are escalating enforcement, and identity-driven breaches continue to expose how fragile access controls remain. Also covered: NIS2 enforcement actions, a 6.9 million record breach tied to a single employee account, and rising risks across AI supply chains and credential exposure. Learn more at https://crestvale.io Support the show