AI-Related Activity Is 0.43% of SOC Alerts, Mostly Noise

Summary

An Intezer analysis of enterprise SOC alerts found that AI-related activity made up 0.43% of alerts; 94.1% were noise, 5.8% risks and 0.02% attacks. It recommends tuning detections and restricting AI tools.

Key points

The analysis covered roughly 16.9 million SOC alerts; about 73,000 were AI-related, and that alert share grew 685% from February to June 2026.
Intezer classified 94.1% of AI-related alerts as noise, 5.8% as genuine security risks and 0.02% as real attacks.
The analysis found that confirmed attacks in this alert class were phishing campaigns using AI brand names as lures; it found no confirmed compromises caused by an organization’s own AI agents.
Unsafe agent behavior included running with permission safeguards bypassed, opening a public reverse tunnel, and dumping a macOS keychain to retrieve a token.
Legacy detections often flagged legitimate coding-agent activity as severe threats; the article reports benign rates of 77% to 99% across several noisy detections.
The article recommends tuning detections for routine AI activity, hunting for risky agent behavior and OAuth grants, and isolating AI tools in restricted environments.

Article Details

Publisher: Intezer
Report Period: February-June 2026 (stable reporting window); broader review covered the preceding year
Scope: AI-related SOC alerts and activity across enterprise environments
Sample Size: Approximately 16.9 million SOC alerts, including about 73,000 AI-related alerts; exact number of environments not disclosed
Key Statistics: AI-related alerts were about 0.43% of approximately 16.9 million reviewed SOC alerts.
AI-related alert volume increased 685% from February to June 2026.
The reviewed AI-related alerts were classified as 94.1% noise, 5.8% genuine security risks, and 0.02% real attacks.
Among AI-related alerts, 79.8% received a benign verdict, 81.7% were automatically suppressed, and 5.4% were escalated to a human analyst.
A single detection at one customer accounted for 55% of critical verdict alerts flagging Expand.exe as a lateral-tool-transfer.
Recommendations: Tune legacy detections that generate high-severity alerts on routine AI-agent activity.
Define policies for information shared with third-party AI platforms.
Proactively hunt for permission-bypass flags, unauthorized tunnels, and risky OAuth grants.
Run AI tools in isolated environments, such as Docker containers or virtual machines, with restricted access to credentials and sensitive information.

Indicators of compromise

Type	Indicator	Context
DOMAIN	`gemini-advertisers[.]com`	Identified as suspicious infrastructure used in a fake Google/Gemini Ads invitation lure intended to direct recipients to a malicious site.