Wired Tests 'Instinct' AI Agent: Saves $550, Burns $64, Raises Security Questions

A Wired reporter's hands-on test of an AI agent called Instinct produced a mixed ledger: the tool saved $550, booked restaurant reservations on request, and flagged a phishing scam before the writer fell for it — but it also burned $64 on wasted actions and left unresolved doubts about whether handing an agent that much access is safe. The account, published this week, is one of the few public stress tests of a consumer-facing agent that actually spends money and acts on a user's behalf rather than just drafting text.

The dollar figures are the headline. The $550 in savings came from agent-driven decisions the reporter credited to the tool, while the $64 loss stemmed from actions that went nowhere or were misdirected. That roughly 8.6-to-1 ratio is the kind of concrete data point the agent industry has mostly lacked, since vendors tend to demo scripted tasks instead of publishing net outcomes from open-ended use.

The phishing catch matters as much as the money. An agent that reads email, watches browsing behavior, and intervenes before a user clicks a malicious link is performing a security function that most antivirus tools handle at the network layer. Instinct did it as a side effect of trying to be useful, which suggests agent builders may be sitting on a detection capability they never set out to sell.

The security worry cuts both ways. An assistant with enough access to book reservations, move money, and read incoming messages is also a single point of failure — if the agent is compromised, phished, or simply wrong, the blast radius covers everything the user has connected. The Wired piece does not resolve whether Instinct's safeguards are adequate, and the reporter's own conclusion leans toward 'worth the risk' rather than 'proven safe.'

That tension is now the central design problem for the agent wave. Companies from OpenAI to Anthropic to a long tail of startups are shipping tools that take actions, not just generate answers, and each one has to decide how much confirmation to require, how to log failures, and who eats the cost when the agent gets it wrong. The $64 loss in this test is small, but it is the kind of number that becomes a lawsuit or a refund policy at scale.

What to watch next is whether Instinct's maker publishes its own error rates, transaction limits, and incident data, or lets a single reporter's experience stand as the public record. Until agent vendors report net savings and losses the way banks report fraud, every buyer is running their own experiment — and paying for the mistakes.