Researchers document prompt injections aimed at AI agents on live websites
Instructions written for AI agents rather than human readers are being planted on public websites, including attempts to initiate payments and deletions.
- What happened
- Observation of attempted exploitation in the wild; neither report confirmed a successful compromise. In March 2026, Unit 42 reported hidden instructions on live web pages aimed at AI agents, including an attempt to get an AI-based ad-review system to approve a scam product advertisement, and payloads attempting to trigger Stripe and PayPal payments, database deletion and system-prompt leakage; it catalogued 22 concealment techniques and found 85.2% of cases used social-engineering framing. In April 2026 Google reported scanning Common Crawl (2-3 billion pages per month) and a 32% relative increase in malicious injections between November 2025 and February 2026, including attempted data exfiltration and file deletion, while noting attackers had not yet productionised such attacks at scale.
- Response
- Not applicable (threat research). Unit 42 stated it was not aware of confirmed successful attacks against deployed ad-checking agents; Google documented no confirmed successful compromises.
- Discovered by
- Unit 42 (Beliz Kaleli, Shehroze Farooqi, Oleksii Starov, Nabeel Mohamed) through web telemetry; Google Threat Intelligence (Thomas Brunner, Yu-Han Liu, Moni Pande) through Common Crawl scanning; Forcepoint X-Labs published related payload analysis.
- Who was affected
- Operators and users of AI agents that read public web content
- Patterns
- Sources
- Fooling AI Agents: Web-Based Indirect Prompt Injection Observed in the WildPalo Alto Networks Unit 42, 3 Mar 2026
- AI threats in the wild: The current state of prompt injections on the webGoogle, 23 Apr 2026
- Indirect prompt injection is taking hold in the wildHelp Net Security, 24 Apr 2026
Related
Incidents with the same pattern
What does this ask of people?
Most entries in the record are not about machines turning against people. They are about delegation: a task, an inbox, a key, a terminal, handed to a system that did more, less or something else. Futures literacy starts with noticing that.