Agent-to-agent coordination is dangerously exposed to prompt injection, goal hijacking, and unguarded transactions

Summary

As users let their agents talk to other people's agents to negotiate schedules, terms, and logistics, they inherit severe, unsolved security and reliability risks: prompt injection and goal hijacking can redirect an agent mid-negotiation, memory is stored as unvalidated free text, and agent workflows skip the input validation, process isolation, and access control that traditional software relies on. The same trust gap shows up when agents execute anything consequential (like a payment) — timeouts and retries can double-charge, and there is no clear recovery from a halfway failure. With no trustworthy guardrails, handing real negotiation and transaction authority to an agent that converses with untrusted counterparties is a liability no founder, sales rep, or recruiter can accept — and no one is solving it for the cross-party coordination use case.