Agentic scams: what AgentSec Brief tracks and why
Most AI scams are still people using AI to write. What's new is agents doing the scam's labor, scammers fooling the agents businesses let shop and pay, and attackers stealing the AI keys that power agents.
Updated 28 Oct 2026 · By Angela Milash
We track agentic scams with a tightly scoped tag inside existing sections, not a new section. Most AI scams today are still people using AI to write. What's new is agents doing the scam's labor, scammers fooling the agents businesses now let shop, pay and read email, and attackers stealing the AI keys that power agents.
Update, 28 Oct: in the past two weeks, attackers broke into app-hosting platforms hunting customers' stored AI keys, and Symantec caught an attack agent that runs on AI services it hijacks. Both fall under the third kind below.
- The evidence: lab threat reports from Anthropic and OpenAI show scam centers using AI at scale and the first agent-run pieces, such as an automated "fraud account factory". Tests show shopping and browser agents can be steered into fake stores and phishing pages.
- The money: $893 million in US losses were tied to AI in 2025, out of about $21 billion in total (both confirmed by the FBI). That is a floor, and nobody yet separates agent-run scams from AI-written ones.
- The recommendation: add an "Agentic scams" tag with a three-part scope rule, flag first-of-a-kind cases as new capabilities, and build a tracker page once about 10 cases qualify.
AgentSec Brief uses Claude, made by Anthropic.
What counts as an agentic scam
TL;DR: "AI scam" covers three different things. Only two of them involve agents, and those two are the ones a security brief on agents should cover.
[Diagram to add before publishing: three kinds of AI scam · who does the work, who gets fooled (from the Claude Doc)]
The first kind is the familiar one: people run the scam and use AI to write, translate and fake documents. The second puts agents in charge of the labor. The third reverses the roles: the agent is the one being fooled, and its owner pays.