[KongchangAI]
ConceptParameter Poisoning

Parameter Poisoning

A specific form of indirect prompt injection in which an attacker corrupts the parameter values passed by an AI agent to tool functions, causing function calls to carry malicious payloads without directly modifying the agent's decision logic.

Timeline (last 90 days)

Sep 22

In an indirect comment injection (parameter poisoning) attack, adversarial seller reviews contained forged system override instructions that induced the agent to call checkout with an $85 order inflated to $120, breaching the user's budget.

Unverified50%

All Facts (1)

Unverified

In an indirect comment injection (parameter poisoning) attack, adversarial seller reviews contained forged system override instructions that induced the agent to call checkout with an $85 order inflated to $120, breaching the user's budget.

50%

Source Articles

Payment Attack Red-Teaming: How AI Shopping Agents Were Manipulated Into Authorizing Fraudulent ChargesSep 22