Blast Radius

Blast radius is the extent of the damage a single failure, mistake or compromised credential can cause before something stops it. It is a way of asking not how likely is this to go wrong but how far does it get when it does, and the second question tends to be the more useful one, because the probability of a mistake over a long enough period approaches certainty while the reach of that mistake is something you can design. The measure applies to whatever can fail. A leaked API key with organisation-wide scope reaches every workspace; the same key scoped to one project reaches one project. A migration that runs against a shared database touches every tenant; the same migration against partitioned data touches one. An automation with permission to send email as anyone can embarrass the company; the same automation restricted to one mailbox cannot. In each pair the likelihood of the failure is identical and the consequence differs by orders of magnitude. The levers are unglamorous and well understood. Scope credentials narrowly and give them short lives, so a leak is bounded in both what it reaches and how long it lasts. Separate environments and tenants so a fault cannot travel sideways. Put limits on the things that consume resources, since an unbounded loop is a self-inflicted outage. Roll changes out gradually, because a bad deployment that reached a small share of traffic is an incident and one that reached everything is an outage. Make the destructive operations require a second party or a deliberate step, since the difference between a recoverable and an unrecoverable mistake is often one confirmation. AI systems changed the shape of this in a specific way. An assistant that only answers has a blast radius bounded by what it can read, which is already larger than most teams assume once a retrieval index has crawled everything a service account could see. An agent that acts has a blast radius bounded by its permissions and its ability to loop, and it can reach that bound faster than a person would. Broad write access granted for convenience during a prototype is the single most common way a small AI project acquires a large blast radius, and prototypes are exactly where such grants are made and forgotten. There is a cost to containment, and pretending otherwise leads to designs nobody follows. Narrow scopes mean more credentials to manage, isolation means more infrastructure, staged rollouts mean slower releases. The point is to spend that cost where the consequence is large — on the actions that move money, delete data, message customers or cross a tenant boundary — rather than uniformly everywhere, which is how containment programmes become paperwork and stop being protection.