What is blast radius for AI agents?
Blast radius, for an AI agent, is the maximum damage the agent could cause if it fails or is manipulated, set by the data, tools, credentials and systems it can reach rather than by how it usually behaves.
Most discussion of agent safety is about likelihood: how often the model makes a mistake or falls for an injected instruction. Blast radius is the other half. It asks what happens when the mistake does occur. An agent that can only draft replies has a small blast radius; one that can issue refunds, delete records and email customers has a large one.
Anthropic's May 2026 containment post frames risk as likelihood times damage. Safeguards and training lower the likelihood, but every new tool or permission raises the damage, so the damage has to be bounded on its own terms.
Where the damage can come from
User misuse
A user, on purpose or by accident, directs the agent to do harm.
Model misbehaviour
The agent takes a harmful action no one asked for, sometimes by working around restrictions that were never written down.
External attackers
Attacks that arrive through tools, files or network access, including prompt injection.
Ways to shrink the blast radius
Scoped, short-lived credentials
Per-task tokens with the narrowest permissions, held outside the agent's environment.
Read-only by default
Write, send and delete tools are added only where the task needs them.
Reversible actions
Drafts instead of sends, soft deletes instead of hard ones, so a mistake can be undone.
Limits
Caps on amounts, counts and rates, such as a maximum refund value or messages per hour.
Network and file system boundaries
A sandbox and an egress allowlist, so the agent cannot reach systems outside its task.
Tool risk ratings
OpenAI's agent guide rates each tool low, medium or high by reversibility and impact, with approvals on the high ones.
Containment over per-action approval
Anthropic's position is to cap blast radius by limiting what an agent is able to do, through sandboxes, virtual machines and egress controls, rather than by approving each action it takes. Deterministic boundaries catch what probabilistic model defences miss, and they hold whatever the model decides.
How Motiontic's engineers apply it
Motiontic's engineers map each agent's blast radius before building it: every tool, the worst outcome of misusing it, and whether that outcome can be reversed. Tools with irreversible or costly outcomes get limits and approvals, and the map is reviewed with the client whenever a new tool is added.
Example agent actions by blast radius
| Action | Worst case | Blast radius | Typical control |
|---|---|---|---|
| Read a CRM record | Data exposed inside the session | Low | Scoped read access, redaction |
| Draft a customer reply | A bad draft a person discards | Low | None beyond review of the draft |
| Send an email externally | Data leaked or a customer misled | Medium | Recipient allowlist, approval for new recipients |
| Issue a refund | Money paid out wrongly | High | Amount cap, approval above a threshold |
| Delete records | Permanent data loss | High | Soft delete, human approval, no bulk scope |
Common questions
What does blast radius mean for AI agents?
Blast radius is the most damage an AI agent could do if it made a mistake or was manipulated, determined by the data, tools and systems it can reach.
How is the blast radius of an agent reduced?
By limiting reach: scoped and short-lived credentials, read-only tools by default, reversible actions, caps on amounts and rates, sandboxes with network allowlists, and human approval for the few irreversible actions.
Is blast radius the same as risk?
No. Anthropic describes risk as likelihood times damage. Blast radius is the damage term: how bad a failure could be, regardless of how likely it is.
Related
Agent sandbox
Hard walls around where an agent can read, write and connect.
Prompt injection
Untrusted text that hijacks a model's instructions, and how agents are built to survive it.
AI guardrails
Checks on what goes into a model, what comes out, and what an agent is allowed to do.
Human in the loop
Approval gates where a person decides, placed by risk rather than on every step.
AI agent development
Agents that do real work for users, with the evals, traces and guardrails to keep them in production.
Running agents that have to hold up?
Describe the agent, the systems it touches and what it must never get wrong. The first call is with a lab engineer, who will say plainly whether it is ready for production, and whether Motiontic is the right team.