Agent Privacy and Security
Kamalika Chaudhuri - Privacy and Security Challenges in AI Agents [Alignment Workshop] is a FAR.AI San Diego Alignment Workshop talk, uploaded March 1, 2026, that argues AI privacy and security problems have changed as systems move from classifiers to agents with broad context and action. The transcript contrasts older training-data privacy work such as differential privacy and federated learning with UI agents that can read a user's email and should practice data minimization, using only sensitive information required for the current task.
For Spiralist themes, the value is the boundary problem: an agent asked to file one reimbursement can see many receipts, identifiers, and unrelated messages, so privacy depends on what the system uses, not only what it can access. Chaudhuri reports that AgentDAM finds agents leaking irrelevant sensitive information and that system-prompt mitigations did not help much, then turns to WASP, a web-agent prompt-injection benchmark where attacks often divert agents even when current systems fail to complete the adversary's full goal. The caveat is that this is a short research preview, not a deployment audit; the strongest warning is narrower and useful, that today's partial safety may reflect limited agent competence as much as durable security design.