Last week, we released Auto-review in Codex. Until now, users had two choices:
Default mode, which requires frequent human approval, and Full Access mode
which removes friction at the expense of oversight. Auto-review offers an alternative path. It replaces user
approval at the sandbox boundary with review by a separate agent. Internally, Auto-review increases our confidence
in running long agentic tasks without synchronous human oversight.
In Auto-review mode, Codex sessions stop for human approval roughly 200x less often than in manual approval mode,
while still catching many of the actions humans would want stopped. Most actions run without approval in the
sandbox. For the small fraction that need review, Auto-review approves around 99%. Even when Auto-review rejects
an action, Codex often recovers on its own by finding a safer way to make progress.
Auto-review performed well on our dangerous actions evaluations, giving us enough confidence to deploy it
internally. It is conservative enough for our security teams and unobtrusive enough for our research and deployment
workflows.
This is an unofficial automated linkpost.
Last week, we released Auto-review in Codex. Until now, users had two choices: Default mode, which requires frequent human approval, and Full Access mode which removes friction at the expense of oversight. Auto-review offers an alternative path. It replaces user approval at the sandbox boundary with review by a separate agent. Internally, Auto-review increases our confidence in running long agentic tasks without synchronous human oversight.
In Auto-review mode, Codex sessions stop for human approval roughly 200x less often than in manual approval mode, while still catching many of the actions humans would want stopped. Most actions run without approval in the sandbox. For the small fraction that need review, Auto-review approves around 99%. Even when Auto-review rejects an action, Codex often recovers on its own by finding a safer way to make progress.
Auto-review performed well on our dangerous actions evaluations, giving us enough confidence to deploy it internally. It is conservative enough for our security teams and unobtrusive enough for our research and deployment workflows.
Continue reading at alignment.openai.com →