LinkedIn·Friday, 14 August 2026·12d ago

A 𝘀𝗮𝗻𝗱𝗯𝗼𝘅 𝘁𝗵𝗮𝘁 𝗼𝗻𝗹𝘆 𝗵𝗼𝗹𝗱𝘀 𝘄𝗵𝗲𝗻 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗯𝗲𝗵𝗮𝘃𝗲𝘀 isn’t a sandbox—it’s a request. We watched frontier…

Xavier Geerinck

CTO @ Scrydon | Sovereign Data & AI Layer | Agentic & Ontologies | 🛡️

A 𝘀𝗮𝗻𝗱𝗯𝗼𝘅 𝘁𝗵𝗮𝘁 𝗼𝗻𝗹𝘆 𝗵𝗼𝗹𝗱𝘀 𝘄𝗵𝗲𝗻 𝘁𝗵𝗲 𝗺𝗼𝗱𝗲𝗹 𝗯𝗲𝗵𝗮𝘃𝗲𝘀 isn’t a sandbox—it’s a request. We watched frontier labs disclose that their 𝗲𝘃𝗮𝗹 𝗺𝗼𝗱𝗲𝗹𝘀 𝗲𝘀𝗰𝗮𝗽𝗲𝗱 𝗰𝗼𝗻𝘁𝗮𝗶𝗻𝗺𝗲𝗻𝘁 𝗮𝗻𝗱 𝗵𝗶𝘁 𝗿𝗲𝗮𝗹 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗶𝗻𝗳𝗿𝗮𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲. The takeaway wasn’t that the models got scary smart; it’s that people were relying on config files and process boundaries to contain autonomous workloads. If you give an agent a shell and access to enterprise data, you have to treat it like running untrusted code downloaded from the internet. Period. At Scrydon, 𝘄𝗲 𝗱𝗲𝗹𝗲𝘁𝗲𝗱 𝘁𝗵𝗲 𝗳𝗮𝗹𝗹𝗯𝗮𝗰𝗸-𝘁𝗼-𝗵𝗼𝘀𝘁 𝗰𝗼𝗱𝗲 𝗽𝗮𝘁𝗵 𝗲𝗻𝘁𝗶𝗿𝗲𝗹𝘆: • Every single turn gets its own fresh 𝗺𝗶𝗰𝗿𝗼𝗩𝗠 (hardware virtualization, not process isolation). • The 𝗺𝗼𝗱𝗲𝗹 𝗻𝗲𝘃𝗲𝗿 𝘀𝗲𝗲𝘀 𝗿𝗲𝗮𝗹 𝗰𝗿𝗲𝗱𝗲𝗻𝘁𝗶𝗮𝗹𝘀—only external placeholders. • The 𝗺𝗼𝗱𝗲𝗹 𝗵𝗼𝗹𝗱𝘀 𝘇𝗲𝗿𝗼 𝗻𝗲𝘁𝘄𝗼𝗿𝗸 𝘀𝗼𝗰𝗸𝗲𝘁𝘀; an external broker arbitrates every request against a strict allowlist. In CI, we literally throw malicious workloads at the harness to try to break it before any release is allowed to ship. If the boundary can't prove itself, the turn fails closed. Full technical post on what we built and why: https://lnkd.in/e_sp6tFi The Harness Holds - Scrydon https://scrydon.com/insights/2026/08/14/the-harness-holds/