Forcepoint warns about AI memory poisoning attacks

Security researchers are warning about a persistent threat known as memory poisoning, which allows attackers to inject misleading or malicious memory items into an AI assistant's persistent memory layer. Unlike prompt injection, the effects of which disappear after a session ends, memory poisoning can corrupt an agent's long-term logic and cause it to repeatedly output flawed information or compromise workflows weeks later. Experts suggest that organizations can mitigate the risk by tracking memory provenance, implementing strict access controls, and keeping humans in the loop for irreversible actions. Open article →