Web Search Defaults to Cached Results, Not Live

The mistake

Asking Codex a question that needs genuinely current information and assuming its web search is hitting the live internet by default, the way a browser search would — then getting an answer that's noticeably stale for something that changed recently.

Why this happens

Web search defaults to a cached, OpenAI-maintained index rather than live retrieval, specifically to reduce exposure to prompt injection from untrusted live content — a cached, curated index is a smaller attack surface than fetching and reading arbitrary live pages on demand. That's a deliberate security tradeoff, not an oversight, but it does mean the default behavior isn't "search the current internet," even though that's the most natural assumption for anyone used to how search normally works.

Why it matters

For most questions, the cached index is close enough that the difference never surfaces. For anything genuinely time-sensitive — a very recent event, a page that changed in the last few days, a fast-moving situation — a default cached search can produce an answer that sounds current but isn't, with nothing in the response flagging that it might be working from a slightly stale snapshot rather than the live page.

The fix

If a task genuinely needs live retrieval, force it explicitly rather than relying on the default — live mode is available on request, and it's also the default automatically under the least-restricted sandbox posture, where the injection-risk tradeoff is judged differently since that mode already assumes a higher trust level. For anything where staleness would actually matter, it's worth checking explicitly which mode a session used rather than assuming the answer reflects the current state of the page it's citing.

Worth being specific about what "stale" means here, too — a cached index isn't wrong the way a broken link is wrong, it's a snapshot from some earlier point that may simply predate whatever changed. That distinction matters when deciding how much to trust a search result for a task where the answer might have shifted recently — treating a plausible-sounding cached answer as current is the actual risk, not the caching itself.

See also

The sandbox and approval modes reference covers how the injection-risk reasoning behind this default connects to the broader posture Codex takes under different sandbox settings.