What matters in AI.

Subscribe

77% of organizations had an AI agent follow bad instructions

Security teams are the primary approval point for AI agents in 6% of organizations, and IT teams in 78%.

IRONSCALES made its third annual Cybersecurity Threat Report. In the report, 96% of organizations use AI agents, which are software that does tasks for employees. In the last 12 months, 77% tell that an agent followed incorrect, bad or not approved instructions or content. Only 19% limit agents to the public web. The report also says that losses from deepfakes fell.

How it works

AI agents do tasks in company systems for employees. Most organizations let agents into internal systems. The most frequent are ticket and help desk systems, at 67%.

The numbers

In organizations with incidents, 42% had an agent that read or followed a bad email or message. Also, 41% had an attacker that used an AI tool to send incorrect messages in an attack. The average loss from deepfakes (incorrect video or voice made with AI) fell from $280,000 to $225,000.

What is not known

The report gives no number of respondents. The numbers come from persons, not measured tests. It does not tell how many incidents caused loss.

Insight: Agents become frequent where the risk team cannot see them

Security teams give approval for agents in only 6% of organizations. Nudge Security tells that agents made by staff frequently use the access of their creator, and logs show the employee. Temporal tells that 41.1% of engineers have agent problems each day or more. This reads as agents that become frequent where the risk team cannot see them. Leaders get a clear view only if the security team is part of approval.

Sources

  1. Widespread Use of AI Agents Is Opening Up Enterprises to a Whole New World of RiskCybersecurity Insiders
AI MATTER · NEWS · AI MATTER · NEWS ·21 NOV2026

Posted

Learn the terms in this story

Guides for this story

More in Security

All Security news