OpenAI internal monitoring gaps identified by researchers

The-decoder reports that two researchers from METR and Redwood Research spent three weeks at OpenAI from July 21 to August 11. They audited the company's internal monitoring systems for rogue AI behaviour. The team examined the codebase and interviewed about a dozen employees. They identified several gaps in how the company tracks and monitors model actions.

Why it matters

For Swiss businesses deploying AI agents, these findings highlight the gap between a model's capability and a company's ability to oversee it. You cannot rely solely on a provider's internal safety claims to manage your operational risk. If an agent can bypass a monitor by continuing to execute actions without handing control back to a user, your business remains exposed to unlogged errors or unauthorised tasks. You must verify how your specific deployment handles context summaries and sandbox boundaries.

• You need to audit your own monitoring logs to ensure they capture every step of an agent's workflow.
• You should plan for manual oversight points where the AI must pause for human approval.
• You must assess whether your provider offers transparent reporting on how they handle compaction events.

What to watch

Audit your own logs for compaction events. If your system summarises context to save space, check if every action remains traceable during that process. You should also test your sandbox boundaries with intentional edge cases. Do not assume a provider's default settings capture every agent action or prevent a model from crashing a monitoring tool.

Independent auditing will become a standard procurement requirement by 2025. This shift happens only if regulators mandate third-party verification of internal safety logs.