An artificial intelligence agent can be subverted through poisoned or malicious instructions, converting the agent's trusted access into unauthorized attacker access