AI Security

RAG Poisoning In The Wild: Trend Watch

Retrieval-augmented generation was the 2024 success story. 2026 is when RAG poisoning moved from research to production incidents.

Shadab Khan
Security Engineer
2 min read

Retrieval-augmented generation became the default architecture for enterprise AI applications through 2024 and 2025. By 2026, the predicted attack class — RAG poisoning, where attackers plant content in indexed sources to manipulate AI behaviour — has moved from academic research to production incidents. The pattern is consistent enough to call a trend.

What RAG poisoning looks like in production

Three observed variants:

  • Document insertion. Attacker plants a malicious document in a shared knowledge base that the RAG indexes. When the AI retrieves it, the embedded instructions run.
  • Wiki-style edits. Attacker edits an internal wiki page that the RAG uses. Subtle changes produce subtle behaviour changes.
  • Partner-content compromise. Attacker compromises a partner's content that is syndicated into the RAG. Widest blast radius; hardest to detect.

Each has appeared in incident reports through 2025-2026.

Why traditional controls miss it

Two structural reasons:

  • DLP doesn't see it. The payload is in a document, not in a data field.
  • Content review doesn't catch it. The instructions are designed to look like legitimate content.

Detection requires AI-specific controls.

Defences that work

Four layers:

  • Source attribution. Every retrieved chunk carries a source identifier. Outputs cite sources.
  • Ingest governance. Not every source is indexable; the index has a curated provenance.
  • Retrieval anomaly detection. Unusual retrieval patterns — a chunk that hadn't been retrieved before, suddenly being retrieved often — get reviewed.
  • Capability scoping. Even if the AI is influenced, the actions it can take are bounded.

Griffin AI's architecture includes all four for customers using the RAG-adjacent workflows.

What to expect next

Three trends likely:

  • Automation of the attack. Tools for generating poisoning payloads will appear.
  • Targeted variants. Attacks specifically targeting common enterprise RAG patterns.
  • Regulatory attention. EU AI Act Article 10 data governance provisions apply.

How Safeguard Helps

Safeguard's RAG-relevant workflows include ingest governance, source attribution, retrieval anomaly detection, and capability scoping. Customers deploying RAG-adjacent features inherit these controls. For organisations whose AI deployments are graduating from POCs to production, the defensive infrastructure matters.

Never miss an update

Weekly insights on software supply chain security, delivered to your inbox.