AI Agent Blast Radius: Scope the Damage First
AI agent blast radius replaces is this agent safe with four answerable questions: reach, reversibility, audience and rate, scored on a worked example.
An AI agent's blast radius is the full set of things that go wrong if the agent does the worst thing it is capable of doing. Not the likely outcome, the maximum one. The term is borrowed from site reliability work, where the question is always how far a failure spreads before something stops it. Applied to agents it replaces an unanswerable question, "is this agent safe", with four questions that have concrete answers and let you compare two setups honestly. Once you have measured the radius, the usual way to shrink it is applying the principle of least privilege.
Why "is it safe" is the wrong question
Agent safety discussions stall because the risk gets framed around intent. Will the model do something malicious? Could it be tricked? Those are real questions with no stable answer, because model behaviour changes with every release and every prompt injection technique.
Blast radius sidesteps them. It asks what the agent's permissions make possible, which is a property of your configuration rather than the model's disposition. A read-only agent with a bad day writes nothing. An agent with production database credentials and a good day is still one confused instruction from a problem. The permissions are yours to set, so that is where the leverage is.
This is the principle of least privilege applied to a new kind of user. NIST's glossary defines it as a security principle that a system should restrict the access privileges of users, or processes acting on their behalf, to the minimum necessary to accomplish assigned tasks. Grant what the task needs, and nothing held in reserve for convenience.
The four questions
Score each one. Any of them scoring high is a finding, not a detail.
1. Reach. What can it touch?
Enumerate it literally: which systems, which accounts, which files, which API keys are in scope. The answer is not what the agent is instructed to use, it is what its credentials permit. An agent told to read one table, holding a key that grants the whole schema, has the reach of the key.
2. Reversibility. Can you undo it?
Sort every action the agent can take into three buckets. Reversible in seconds, a local file edit under version control. Reversible with work, a database write with a recent backup. Not reversible, a sent email, a published post, a transferred payment, a deleted account. The third bucket is where blast radius concentrates, and irreversible AI agent actions covers it on its own.
3. Audience. Who sees the mistake?
An error in a draft is invisible. An error in a customer-facing reply is a support problem. An error posted publicly under your company name is a reputation problem that outlives the fix. The same wrong output has wildly different costs depending on who receives it, and audience is often the cheapest dimension to shrink: add a review step and the audience becomes one person.
4. Rate. How fast, and how many times?
A single wrong action is an incident. Four hundred wrong actions before anyone notices is a different event. Rate turns a small error into a large one with no change in the error itself. An agent running unattended overnight has a far larger blast radius than the same agent run interactively, purely because nothing interrupts it.
A worked example
A support agent with inbox access, asked to draft replies to incoming tickets. Two configurations, same model, same prompt.
Config A | Config B | |
|---|---|---|
Reach | Full mailbox, send permission | Single folder, read only, writes to a drafts queue |
Reversibility | Sent mail cannot be recalled | Drafts deleted in one click |
Audience | Customers, directly | One support agent |
Rate | Unattended, all night | Processes on demand, batch of 20 |
Worst case | 300 wrong answers sent to customers under your name, overnight, unnoticed until morning | 20 bad drafts a human reads and deletes |
Config A's worst case is a business incident. Config B's is twenty minutes wasted. The gap is not model quality or prompt engineering. It is four configuration choices, each of which costs almost nothing to make differently at the start and a great deal to retrofit after the incident.
Note what did not change: the agent is just as likely to be wrong in B as in A. Blast radius work does not reduce the error rate. It reduces what an error costs, which is the part you control.
Shrinking each dimension
Reach. Give the agent its own credentials, never a human's. Scope them to the narrowest role that completes the task. Read-only wherever a write is not strictly required. Limiting what an agent can reach on your network covers the network layer, and sandboxing an agent the execution layer.
Reversibility. Put irreversible actions behind a confirmation. A queue the agent writes to and a human releases converts most of them into reversible ones.
Audience. Default to drafts. The review step is the single highest-value control available, and it is covered as a pattern in what human in the loop means in AI.
Rate. Cap actions per run and per hour. Where money is involved, cap the amount too, as in setting spending limits for AI agents.
When a large blast radius is acceptable
Sometimes it is, and pretending otherwise leads to theatre rather than security. A large radius is defensible when the work genuinely requires it, the state is cheaply recoverable, and someone is watching. An agent with broad write access to a scratch environment that is rebuilt from scratch every morning has enormous reach and near-total reversibility, which is a reasonable trade.
What is not defensible is a large radius that nobody has measured. The four questions take about ten minutes per agent. Writing the answers down means the next person can see what was decided, and it gives you something to compare when you add the next capability. For the broader landscape this sits inside, AI risks is the overview.
FAQ
What does blast radius mean for an AI agent?
The complete set of consequences if the agent takes the most damaging action its permissions allow. It measures what is possible given the configuration, not what is likely given the model.
How is blast radius different from a risk assessment?
A risk assessment multiplies likelihood by impact, which requires estimating how often a model will misbehave. Blast radius deliberately ignores likelihood and measures only the maximum impact, because that part you can calculate precisely from the permissions.
What is the fastest way to reduce an agent's blast radius?
Make its output a draft instead of an action. That single change shrinks audience to one person and makes nearly every action reversible, usually without reducing the agent's usefulness much.
Should I score blast radius before or after building the agent?
Before, because reach and reversibility are decided when you choose the credentials and the integration, and both are awkward to change later. Re-score whenever you add a permission or a tool.
How did this land?
About the author

Senior Editor, AI & Product
Cecilia leads the Swarmz editorial desk. She has spent a decade turning complex AI and product topics into writing people actually finish, and she owns the blog's quality bar.


