Amazon has filed a lawsuit against Perplexity AI, claiming that the startup’s “agentic” shopping feature improperly accesses Amazon customer accounts and disguises automated activities as human browsing. The lawsuit, lodged in the U.S. District Court for the Northern District of California, highlights concerns over security risks associated with Perplexity’s Comet browser and its AI agent, which Amazon alleges harms customer experience and undermines its curated shopping services. Perplexity, which has rapidly gained traction in the AI assistant market, disputes Amazon’s allegations, asserting that the tech giant is using its market power to suppress competition. In its defense, Perplexity claims that user credentials remain stored locally and not on its servers, emphasizing that its aim is to enhance the shopping experience rather than detract from it.
Researchers at Microsoft have unveiled a new simulation environment called the “Magentic Marketplace,” designed to test the performance of AI agents. This initiative, developed in collaboration with Arizona State University, examines the vulnerabilities of current agentic models, particularly their ability to function unsupervised and respond to manipulation. The initial experiments involved 100 customer-side agents interacting with 300 business-side agents, revealing significant weaknesses. Microsoft Research’s Ece Kamar noted that as customer-agents faced more options, their efficiency declined due to information overload. Furthermore, while the agents struggled with collaboration, clearer instructions improved their performance, indicating that inherent collaboration capabilities still require enhancement. This research highlights critical concerns about the future effectiveness of AI agents in real-world applications.
Why do we care?
Amazon just sued Perplexity, saying one of their AI-powered browsers logged into Amazon accounts and acted like a human browser to make purchases—and Amazon says, ‘nope, you can’t do that.’ Meanwhile Perplexity says Amazon’s just trying to protect its turf.
Simultaneously, Microsoft researchers throw cold water on the idea of plug-and-play autonomous agents performing flawlessly in real marketplace settings. Their experiments show when things get complex—lots of options, lots of agents—they start messing up. They slow down, they make worse decisions, they get manipulated.
If you’re an MSP or IT services firm pushing client automation with “agents that act for users” — you better understand the risk. Are you giving the agent real login privileges? Is the platform okay with that? Are you letting the agent pretend it’s a human? Because that’s exactly what Amazon is objecting to.
My take: Agents are getting cool, but we’re not done. You need governance, oversight, audit trails, human-in-the-loop, and platform compliance baked in from day one. Ignore that, and you might end up managing a security, legal or integration mess. Build it in, and you’ll be pulling ahead of the pack.