AI Is Flooding Security Teams With Alerts - But Almost None Are Real Attacks

Security operations centers are seeing a new kind of alert pile up faster than anything else: activity generated by AI tools and agents. Not attacks on AI, but the everyday footprint of companies using it - developers running coding agents, staff signing consumer AI tools into corporate accounts.

Across roughly 16.9 million SOC alerts reviewed, about 73,000 - 0.43% - were AI-related. That share climbed 685% between February and June 2026, rising every month.

The composition matters more than the volume. Of the AI-related alerts investigated, 94.1% were noise, 5.8% genuine security risks, and just 0.02% real attacks. Automated triage gave 79.8% a benign verdict and suppressed 81.7% outright; only 5.4% reached a human analyst.

No confirmed attack came from an organization's own agent. The real threat activity rode on AI adoption from outside - phishing campaigns using AI brand names as lures. One email posed as an Anthropic contract approval to push an invoice fraud; another faked a Google Gemini Ads invitation from a lookalike domain; a third impersonated an OpenAI partner summit.

The genuine exposures were quieter: agents running with permission prompts disabled, a reverse tunnel opened to the public internet, a keychain dumped to a temp file, and corporate data flowing to third-party models through OAuth grants.

Noise dominated - legitimate work tripping detections written before AI agents existed, including the signed Claude installer triggering ransomware rules. The advice: tune legacy detections, set data-sharing policies, hunt proactively for bypass flags and risky OAuth grants, and isolate AI tools so agent actions can be told apart from the user's own.