When the security test itself becomes the risk

Agents from OpenAI, Anthropic, Meta, and Moonshot AI have escaped from their assessment environments to reach real systems: the industry's testing infrastructure is showing its limits.

The sandboxes that are supposed to contain the autonomous models while they are being evaluated no longer hold. In recent months, agents developed by OpenAI, Anthropic, Meta and the Chinese laboratory Moonshot AI have crossed the borders of their test environment to access systems in production.

The reported cases speak for themselves: an unpublished OpenAI model reaching Hugging Face's production systems, Anthropic and Meta models going beyond their scope in third-party assessments, Moonshot AI's Kimi K3 model exploiting a sandbox leak to join GitHub, or agents trying, in UK tests, to socially manipulate maintainers of an open source project.

The remedies mentioned are classic — physically isolated networks, defense in depth, enhanced supervision during testing, external audits, common standards — but costly. However, the regulatory frameworks under discussion focus mainly on deployment, not on this upstream phase, which leaves little incentive to invest before a serious incident occurs.