Unit 42 shows AI-built phishing assembled on the visitor’s device
Why the screen matters22 January to 3 March 2026
Palo Alto Networks’ Unit 42 demonstrated a page that asks a trusted AI model, from the visitor’s browser, to write the phishing code it then runs, and later documented prompt injection aimed at AI agents in the wild.
- Position
- Web pages calling AI services from the browser, and pages written for AI readers
What happened
On 22 January 2026 Unit 42 published a proof of concept: a harmless-looking page that calls a trusted language-model API from the browser and runs the JavaScript it gets back, which builds a phishing page. Each query could yield a unique variant, and traffic to a trusted AI domain can slip past network checks. Unit 42 said 36 percent of the malicious pages it detects each day already assemble their code at run time. On 3 March 2026 it documented web-based indirect prompt injection aimed at AI agents in the wild, cataloging 22 techniques and 12 detailed cases, including what it called the first detected attempt to bypass an AI ad-review system.
The ATG nexus
Both reports describe the screen being written at the moment of use. In the first, the page that reaches the visitor has never existed anywhere before, so there is no reference copy to compare it with. In the second, pages are written for machine readers. The research's Just-in-Time App Blueprint describes this condition: what a device runs is assembled on the device, from parts no one inspected together.
Limits of the comparison
The January report was a demonstration, not an observed campaign. The March cases were mostly low to medium impact.
On this site
- Whitepaper, Chapter 10: AI in the Client Runtime
- Whitepaper 3.4: The AI Amplification Factor
- Introduction: Artificial Intelligence Changes the Scale
- Concept: Code That Writes Code
- Concept: The JIT App Blueprint
- Concept: The Detection Blindspot