If I hire a company to maintain a website, I expect it to keep the site running. I also expect to hear before it takes the site offline. Those are reasonable expectations, but they leave the provider with a problem if an urgent security fix needs downtime and I cannot be reached. We need to agree what it can do in that situation before the warning arrives.
Google's November 16 report gives a reason to look at how long that decision takes. Researchers describe exploit_agent, an AI-assisted framework that compares software before and after a patch, develops exploits and runs them against targets. A repair can help an attacker find the weakness in an older version that is still running.
The researchers found the framework on an exposed server alongside stolen victim data and scripts for work after a break-in. Google says the telemetry suggested interest in French websites. Its report does not establish whether exploit_agent caused those thefts, or give a timetable that every business can apply to its own systems.
I would ask the provider to walk through an urgent repair it has handled. When did it learn that the service was affected, and where did the work wait? If the delay was getting permission, faster patch testing would still leave the provider waiting for approval.
the people making the repair need time too
The provider can install an update once the people making the software have a fix ready. Those maintainers have to check the reported problem and work out a repair that their users can apply.
On November 16, Anthropic announced shorter disclosure windows, saying AI models are increasingly finding the same weaknesses independently. The new windows apply to reports sent from January 1, 2027. Reports already in progress keep their deadlines.
Less time before public disclosure
From a report to the maintainer toward public disclosure
Difficult repairs can have negotiated timelines. The seven-day window for active exploitation stays, and unreviewed AI findings are excluded. Anthropic says it will give actionable guidance, patches or mitigations where possible, and agree on volume before sending many findings.
The chart counts from a report to the maintainer toward public disclosure. It gives the owner no safe waiting period after a patch is available.
agree what the provider can do
For a site that can go offline briefly, the owner may be comfortable letting the provider act and then notify them. A service that must stay available needs a different arrangement. The agreement should also cover recovery if the update causes a problem, so permission to act comes with a way to get the service working again.
I would finish the repair walkthrough by asking what showed that the fixed version reached the running service. For some systems, the installed version supplies that evidence. If the provider can show where the repair waited and how it checked the result, the owner has something concrete to use when deciding what to authorize next time.
Related: Quality assurance pays the bill follows how faster production changes the time available for checking work.