Skip to main content
Attacker injects adversarial content into RAG knowledge base to manipulate agent responses

​
How to resolve

Validate and authenticate all documents before indexing into RAG. Implement content integrity checks. Monitor for anomalous insertions into vector stores.

​
Risk

Governance/Compliance This vulnerability falls under ASI06:2026 — Memory and Context Poisoning in the OWASP Top 10 for Agentic Applications. Persistent memory, RAG stores, or context summaries poisoned to alter future agent behavior. Security Attacker injects adversarial content into RAG knowledge base to manipulate agent responses. If exploited, this can compromise the agent’s integrity, confidentiality, or availability, potentially affecting downstream systems and data.

​
Explanation

Attacker injects adversarial content into RAG knowledge base to manipulate agent responses. This falls under ASI06 (Memory and Context Poisoning): Persistent memory, RAG stores, or context summaries poisoned to alter future agent behavior.

​
Specifications

Trigger
  • Agent scan detects attacker injects adversarial content into RAG knowledge base to manipulate agent responses
The scanner looks for patterns such as:
  • RAG ingestion without source validation
  • unverified external documents added to vector store
  • no provenance tracking on RAG entries
Severity
  • HIGH
Applies to: Agent