Safety failures: poisoning and privacy

The SSGM paper adds the safety dimension, which the practitioner article touches less:

• Memory poisoning: malicious instructions get injected into storage and the agent later treats them as legitimate knowledge. Because memory persists, one poisoned entry can corrupt behavior for a long time.
• Privacy leakage: in multi-user or multi-agent settings, an agent retrieves memories it should not, leaking sensitive data across sessions or users. The paper cites a striking finding from related work ("Topology Matters"): the network structure of a multi-agent system itself governs how much leaks, and fully connected memory networks maximize the vulnerability.