Convergence Proxy AI — agentic threat model

Utilizes proprietary Large Meta Learning Models (LMLMs) to adapt to dynamic interfaces. This makes it highly susceptible to adversarial prompt injection embedded in third-party web pages it browses, potentially leading to model reprogramming or mis-aligned actions.

Not certain from the listing — utilizes 'long-term memory' and 'learns from user interactions', which implies a persistent data store or vector database. This introduces risks of memory poisoning and unauthorized exfiltration of sensitive user session data.

Orchestrates complex, multi-step web automation workflows. The primary threat is tool misuse and insecure tool integration, where malicious DOM elements or hijacked web interfaces manipulate the agent's browser automation tools to perform unauthorized actions.

Not certain from the listing — likely deployed via cloud infrastructure with browser sandboxing. Threats include container escape from the browser execution environment and the exposure of active session cookies or API secrets used during automation.

Not certain from the listing — mentions learning from user interactions to enhance performance, but lacks explicit details on real-time guardrails, action logging, or drift detection for automated web workflows.

Not certain from the listing — while the acquisition by Salesforce implies future integration with enterprise-grade compliance frameworks, the listing itself does not specify active identity, authorization, or policy controls.

Designed to integrate into the Salesforce Agentforce platform. This ecosystem integration introduces significant multi-agent trust abuse risks, where compromised or rogue agents can trigger cascading failures or unauthorized actions across enterprise systems.