REF: REC-POISONPCP ARCHIVE
AI SECURITY
AI Recommendation Poisoning
When the target shifts from the human to the AI assistant the human trusts.
Instead of manipulating the decision-maker directly, an adversary may attempt to manipulate the information environment used by an AI system trusted by the decision-maker. This connects Personalized Cognitive Pressure directly to AI security.
Attack chain
ATTACKER
↓INFORMATION ENVIRONMENT
↓AI ASSISTANT
↓ANALYST / EXECUTIVE
↓DECISION
Defensive questions
▸ How does the AI retrieve information?
▸ What sources does it trust?
▸ Can its memory be poisoned?
▸ Can fabricated sources enter retrieval?
▸ Can synthetic consensus affect recommendations?
▸ Can manipulated web information influence generated summaries?
▸ How should high-consequence AI recommendations be verified?
KEY FINDINGS
- ▸The attack surface now includes the AI assistant's retrieval corpus and memory.
- ▸Synthetic consensus can bias generated summaries without targeting the human directly.
- ▸High-consequence AI recommendations need the same verification as human-sourced intelligence.
● LAST REVIEWED: 2026-09-30 · REVIEW STATUS: EDITORIAL VERIFICATION · EVIDENCE-BASED