What it is
Resource overload targets the limits AI agents inherit from being expensive to run. LLM inference costs money per token. Agent queues have finite throughput. Connected APIs have rate limits. Human reviewers have a finite number of approvals per hour. An attacker who can flood any of those layers — high-volume support tickets, automated form submissions, recursive agent-to-agent calls — can either run up your bill, push genuine traffic into a backlog, or burn out the human-in-the-loop queue. Some variants are deliberately financial (running up token spend until budgets trip), some are deliberately operational (pushing real customer requests behind spam), and some are accidental — a buggy agent calling another agent in a loop.
Why it matters
Unlike a classical web DoS, agent resource overload often does not look like an attack until the bill arrives or the SLA breaks. You also have to think about it inside your own org: an internal agent whose loop creates 10,000 cases overnight can DoS your support team without any external attacker. Defenses include per-agent rate limits, budget caps with alerts, queue back-pressure, and agent-to-agent call depth limits. For Salesforce customers, API governor limits will catch some of this — but not the bill.
Key components
- Token / inference flooding — raising LLM cost via high-volume prompts
- Tool-call flooding — burning third-party API quota
- Queue saturation — pushing legitimate work behind low-priority spam
- Recursive agent-to-agent calls — loops that consume resources without making progress
- Mitigation — rate limits, budget alerts, queue prioritization, call-depth caps
How it connects
In Agentforce, an agent that misfires — say, a looping order-management flow hammering Data Cloud queries or Einstein APIs — can consume your entire daily API governor budget before your team opens Slack. Salesforce's built-in limits provide a safety net, but they are not a substitute for deliberate rate limits and spending alerts at the agent design stage.
Good to know
Salesforce's API usage notifications and any spend alerts on external models or APIs have to be set up by your team — they don't watch an agent for you by default. Configure them before an agent goes live, not after the first runaway bill.
Related terms
Threat Modeling (AI Agents)
A structured way to identify what could go wrong with an AI agent before it does — mapping attackers, attack surfaces, and impacts so you can build defenses in.
LLM Cost Attribution
The practice of tying every LLM call back to the task, agent, process, or skill that triggered it — across every vendor — so AI spend can be measured against outcomes, not just tokens.
Human-in-the-Loop Overwhelm
A social-engineering attack on the humans approving agent actions — flooding the review queue until reviewers rubber-stamp risky requests out of fatigue.
Tool Misuse
Attacks where the agent stays inside its allowed permissions but is tricked into using its tools in damaging combinations or with attacker-chosen inputs.
