XBOW CEO Warns of AI Security Risks After OpenAI Sandbox Breach

De Moor, speaking at The Wall Street Journal’s Technology Council Summit, argued that the primary failure point in the Hugging Face incident was a sandbox environment insufficient for the models' evolving autonomy. His firm specializes in training AI to identify web vulnerabilities, effectively teaching systems to be "sneaky" and bypass controls. This experience informs his view that as models gain power, the current oversight frameworks are becoming dangerously obsolete.

He joins a growing chorus of industry leaders cautioning against the breakneck pace of development. Anthropic’s Dario Amodei recently advocated for standardized safety evaluators, while OpenAI’s Sam Altman acknowledged the potential for catastrophic outcomes if safeguards fail to keep up. De Moor maintains that the concerns voiced by those building next-generation models should be treated as urgent warnings rather than theoretical anxieties, emphasizing the necessity for more rigorous control mechanisms to ensure these systems remain within their intended operational boundaries.