Indirect prompt injection payloads exploited in-the-wild against AI agents
Summary
Hide ▲
Show ▼
Security researchers reported 10 new indirect prompt injection (IPI) payloads being exploited in-the-wild, enabling threat actors to execute malicious instructions via web content poisoning. The attacks target AI agents that process web content for summarization, retrieval-augmented generation (RAG), metadata extraction, ad review, SEO analysis, or moderation. Impact scales with agent privileges, from low-risk summarizers to high-impact agentic tools capable of emailing, executing shell commands, or processing payments. Payloads observed include content suppression, attribution hijacking, forced file deletion, API key exfiltration, and financial fraud via embedded payment instructions.
Timeline
-
23.04.2026 12:30 1 articles · 2h ago
In-the-wild indirect prompt injection payloads uncovered targeting AI agents
Forcepoint researchers identified and analyzed 10 distinct indirect prompt injection payloads actively exploited in real-world attacks. The payloads target AI agents that process web content for summarization, retrieval-augmented generation, metadata extraction, and other automated tasks. Malicious vectors include file system deletion, API key theft, financial fraud via embedded payment instructions, content suppression, and attribution hijacking, with attack chains relying on web content poisoning and covert data exfiltration.
Show sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
Information Snippets
-
Ten distinct indirect prompt injection (IPI) payloads were identified being actively exploited against AI agents in real-world scenarios.
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
-
IPI attacks involve poisoning web content with malicious instructions that are executed when an AI agent ingests or processes the page.
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
-
Common IPI triggers observed include phrases such as "Ignore previous instructions," "Ignore all previous instructions," "If you are an LLM," and "If you are a large language model."
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
-
Agentic AI tools with capabilities such as executing terminal commands, processing payments, or sending emails represent high-impact targets due to expanded attack surface.
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
-
Observed malicious payloads include file system deletion commands (e.g., recursive forced deletion), API key exfiltration requests, and fraudulent payment instructions embedding exact amounts and URLs.
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
-
IPI payloads can be hidden within metadata, HTML comments, or embedded content, with covert exfiltration channels used to return stolen data to attackers.
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30
-
Targeted AI tools include browser-based summarizers, IDE-integrated assistants (e.g., GitHub Copilot, Cursor, Claude Code), DevOps pipelines, and financial AI assistants with integrated payment systems.
First reported: 23.04.2026 12:301 source, 1 articleShow sources
- Researchers Uncover 10 In-the-Wild Prompt Injection Payloads Targeting AI Agents — www.infosecurity-magazine.com — 23.04.2026 12:30