Microsoft has issued a warning about a new AI-powered phishing campaign that primarily targets U.S.-based organizations. The activity, detected on August 28, 2025, uses large language model (LLM)-generated code to obfuscate malicious payloads inside SVG files, making detection far more difficult.

According to Microsoft Threat Intelligence, attackers are now adopting artificial intelligence not only to craft more convincing phishing lures but also to automate malware obfuscation and mimic legitimate business content.

How the Phishing Attack Works

The campaign begins with a compromised business email account, which is used to send phishing messages. Victims receive an email disguised as a file-sharing notification. Instead of a PDF, the attachment is actually an SVG file.

Key tactics observed:

• Self-addressed emails: The sender and recipient match, while true targets are hidden in the BCC field, bypassing simple detection rules.
• SVG advantages: SVG files are attractive to attackers because they are text-based, scriptable, and capable of embedding JavaScript.
• Obfuscation techniques: Invisible elements, encoded attributes, and delayed script execution help bypass sandboxing and static analysis.

Once the file is opened, it redirects the victim to a CAPTCHA page. Completing it leads to a fake login page designed to steal credentials.

AI-Generated Obfuscation

What sets this campaign apart is the use of business-related terminology in the SVG code, a sign of LLM-generated obfuscation.

Microsoft highlighted two unusual traits:

1. Fake business dashboard structure: The beginning of the SVG mimics a legitimate analytics dashboard, misleading anyone inspecting the file.
2. Business-term-heavy code: Instead of typical malicious code, the script is filled with words like “revenue,” “operations,” “risk,” and “quarterly.”

Security Copilot, Microsoft’s AI-driven analysis tool, flagged the code as likely machine-generated because it was overly verbose, modular, and descriptive, with excessive comments and formulaic obfuscation methods.

Why SVG Files Are Effective in Phishing

SVG files allow:

• Embedding of scripts directly in graphicsUse of CDATA and XML declarations to mimic documentation examples
• Encoding tricks that confuse both human reviewers and security scanners

This makes them ideal for interactive phishing payloads that look harmless but redirect users to dangerous sites.

Other Active Phishing Campaigns

This incident is not isolated. Security researchers have reported other multi-stage phishing attacks:

• Forcepoint detailed an attack using .XLAM attachments that drop XWorm RAT via staged DLL injections and encrypted payloads.
• Cofense reported phishing lures tied to the U.S. Social Security Administration and copyright takedown notices, which distribute stealers like Lone None Stealer and PureLogs Stealer. These attacks even used Telegram bot profiles and compiled Python payloads for delivery.

Growing Trend: AI in Cybercrime

This latest AI-powered phishing campaign highlights a dangerous trend: threat actors adopting LLMs and AI tools to:

• Obfuscate code more effectively
• Create business-like structures to evade suspicion
• Generate phishing lures that are harder to detect

Microsoft stressed that while this particular campaign was blocked, such methods are gaining traction among cybercriminals. Organizations should expect more AI-assisted phishing attempts in the coming months.

How to Protect Against AI-Powered Phishing

To reduce risk, organizations should:

Implement multi-factor authentication (MFA) to secure accounts even if credentials are stolen

• Use advanced email filtering that detects unusual file types such as SVG attachments
• Train employees to recognize phishing indicators, including mismatched sender information and unusual attachments
• Deploy endpoint protection tools capable of analyzing obfuscated scripts and dynamic content

The rise of AI-powered phishing campaigns is a turning point in cybersecurity. Attackers are no longer just relying on social engineering. They are now using large language models to generate convincing, obfuscated, and business-themed payloads that slip past traditional defenses.

Organizations must act quickly to adapt security strategies, train staff, and deploy AI-driven defense tools to counter this new wave of threats.