The Pulse
Island Extends Browser Threat Intelligence to AI Agents
Island says its browser threat intelligence evaluates page behavior on the device, whether a person or an AI agent opens the page. Its system combines those verdicts with threat feeds and uses AI agents to review detections and propose fixe
AI.info Team · Written with AI · fact-checked ·
Island AI Threat Intelligence applies the same page-level checks when an employee browses and when an AI agent opens a page on the employee’s behalf. The company contrasts that approach with reputation systems, which classify domains or score metadata and can miss a threat on a trusted site. Island announced the product in a blog post published November 3, 2026.
Why a clean domain can still host an attack
In May 2026, researchers at XLab found attackers had broken into more than 700 sites running Ghost, a publishing platform, through a database injection flaw disclosed in February. The attackers altered articles to show a fake human-verification check; visitors who followed its steps pasted a command that installed malware. By mid-May, antivirus engines had zero detections for the final payload, Island says.
The compromised sites kept their domains and certificates, leaving reputation feeds without a reason to flag them. Island argues that domain histories and other metadata do not show what a page actually does. Its Dynamic Phishing Protection instead inspects a page as it renders, including its structure, layout and scripts.
Island looks for behavior on the first visit
The company says the same analysis can identify a Microsoft-styled sign-in form served from a domain Microsoft does not own, or a reverse proxy that relays traffic to a real sign-in page and keeps the session cookie. Island names Evilginx and Modlishka as examples of adversary-in-the-middle kits. It also looks for scripts that write commands to a user’s clipboard.
For ClickFix attacks, Island watches for a disguised shell command headed for a paste into CMD, PowerShell or Terminal, and can warn or block before the command runs. Administrators choose how far clipboard screening goes, and detections are audited. Island also says it restricts and audits high-risk browser capabilities according to the risk of the destination.
Island describes its verdicts as a combination of page and request behavior, existing threat information, and the identity and device posture of whoever is browsing. Local indicator lists, external feeds, the company’s own research and past verdicts contribute to the decision. High-confidence benign pages load locally, high-confidence malicious pages are blocked, and uncertain pages open in remote browser isolation.
AI agents enter the same security loop
Island says the system applies the same analysis when an AI agent loads a page for an employee. Each consequential verdict becomes a new indicator of compromise and reaches the company’s browser, desktop and network control points. The blog says that a verdict made in one place can therefore inform decisions at the others.
After a verdict, Island AI Investigators, which the company calls Guardian Agents, re-examine verdicts and sessions to look for missed detections and false positives. They diagnose root causes and propose fixes such as a policy change, a domain block or a user notification. Administrators can see issues in Island Insights and receive notifications through webhook or Slack, or pull issues into other tools through an API.
Island cites figures that describe the pressure on reputation-based defenses: Microsoft’s Digital Defense Report 2025 found AI-written phishing had a 54% click-through rate, compared with 12% for lures written by people. The report also found ClickFix accounted for 47% of initial access in incidents Microsoft Defender Experts could trace to a starting point. Island’s own scan of 33,563 public Model Context Protocol servers found that more than a third had a high- or critical-severity finding.
“Threat intelligence spent 20 years asking whether anyone had seen a threat before. The threats that matter now are the ones nobody has seen yet.”
Yogev Shaban, Product Manager (Web Security, Endpoint Security, Threat Intelligence), Island
For the security operations center, the work moves from hunting for context to approving a fix, Island says. Its stated approach puts the page’s behavior, the browsing identity and device posture into the same decision, whether the session belongs to a person or an AI agent.