How DNV unlocks more unattended AI work with Watcher

Why DNV saw a need for monitoring

As agents get more capable, their ability to cause harmful, expensive incidents rises. Reports such as OpenAI’s report on the Hugging Face incident describe how models can harmfully bypass controls, even when harnesses are configured correctly. This increase in risk blocked DNV from letting agents do more work autonomously. To achieve its ambition of expanding agent use, DNV required new security controls to keep up with the threats posed by powerful AI agents.

More autonomy with the same tools

Using GitHub Copilot in manual approval mode made it hard for engineers to step away from the screen. Autopilot mode removes those approvals, so longer unattended runs need other safeguards. Even with sandboxing to mitigate risk, DNV needed oversight of how agents made decisions, especially when sandboxes still allowed agents to work with cloud infrastructure. Independent monitoring adds that oversight alongside existing controls.

After DNV reached out to us for help, we integrated Apollo Watcher into its existing stack. Watcher checks supported tool calls against the team's policies before they execute. It can block dangerous actions, such as exposing credentials or disrupting infrastructure, and ask an engineer to review actions that need human judgment.

As the agent works, Watcher also reviews the wider session in the background. This monitoring looks for risks that emerge across several actions, including attempts to expand access or weaken security controls, and flags findings to DNV's security team. Together, these checks support longer agent runs without requiring an engineer to watch every step or switch tools.

Security without the slowdown

Security teams often worry that new guardrails will slow their engineers down. For Rezoan Alom, an AI Engineer at DNV, Watcher had the opposite effect.

Rezoan continuously pushes the boundaries and gives models harder tasks as they improve. Before Watcher, when something went wrong, he had to build up trust again. With Watcher in the background, he felt more confident delegating longer work, including setting up Azure infrastructure.

He rarely has to think about it. Watcher steps in about once a week, and he makes the call directly in the tool he is already using. In an interview, he described a change in his broader coding-agent use:

He credits both more capable models and greater trust from monitoring.