Blast Radius

Blast radius is the practical measure of how far one failure can spread inside a system before a control meaningfully contains it. In security engineering, the term answers a simple question: if this component is compromised, how much else moves with it?

In agentic systems, blast radius is rarely limited to one model response. A single poisoned retrieval result can shape shared memory, alter planning, trigger tools, influence user approvals, and affect later sessions if state is durable. That is why blast radius is central to reviewing AI agents, MCP-connected runtimes, and wallet-enabled assistants.

Why Blast Radius Matters

A system can still contain vulnerabilities and remain operationally safe if the impact of any one compromise is tightly bounded. The opposite is also true: even a small upstream flaw becomes dangerous when the architecture allows it to spread across trust tiers, tenants, or execution paths.

For AI systems, blast radius is shaped by:

• whether agents share mutable memory,
• whether tools inherit broad credentials,
• whether retries or fan-out replay unsafe actions,
• whether user-facing approvals are independent of agent-authored summaries,
• and whether high-impact sinks perform their own validation.

Reducing Blast Radius

Typical controls include narrow tool scopes, isolated state stores, tenant separation, rate limits, independent policy checks, and sink-time validation before execution or signing. The design goal is not to guarantee that no component ever fails. It is to guarantee that one failure stays one failure.