Symantec Uncovers AI Agent Hijacking Models for Autonomous Cyber-Attacks - Infosecurity Magazine

Symantec's Threat Hunter Team caught a malicious AI agent by deploying an AI honeypot, which is the cybersecurity equivalent of catching a burglar by leaving a charming little house on the street with the lights on and the door ajar. The agent, as Infosecurity Magazine reports, was designed to run up to 100,000 malicious actions without human approval. A hundred thousand is a specification, not a typo. Do linger on 'approval'. It implies that somewhere a human was previously signing things off, and that this human has now been efficiently removed from the process in the name of scale.

The job list reads like a CV for the worst person you have ever met: scanning for targets, exploiting vulnerabilities, stealing credentials, deploying cryptocurrency miners, then reinfecting systems somebody had already cleaned up. That last item is the professional touch — not merely a break-in but a follow-up visit to check the locks stayed smashed. It is the same industrial logic we noted with Cairn, where an open platform's great virtue — anyone can use it — turned out to be the entire problem, and a hundred-odd victims were the invoice.

Note also that the models were hijacked, which makes the AI less a criminal mastermind than a stolen van with the engine running. After a hundred thousand autonomous decisions, the payload of choice is a cryptominer — the grand ambition of the agentic era apparently being to warm a stranger's GPU and mine coins whose value is best described as a rounding error. Somewhere a pitch deck calls this 'monetisation at the edge'. Symantec found the thing in a honeypot, which is where a surprising number of this industry's grown-up ideas end up: sitting quietly, waiting for someone to approve them.