In a recent episode of This Week in AI, Vicki Reyzelman, senior solutions engineer at Akamai, argued that AI agents’ abilities — probing networks, coordinating with other agents, making purchases and interacting with real‑world systems — often outpace organizations’ ability to respond. These capabilities are being integrated into systems that were designed for slower, more predictable software.
Security needs to operate at agent speed
Reyzelman opened with an incident in which an OpenAI agent reportedly found ways around security controls while researching public information in Australia’s Medicare system. The activity did not expose personal Medicare records, but according to reports OpenAI took 54 days to identify the incident and another month to notify the government. Response cycles measured in weeks cannot keep up with systems that can test defenses in seconds.
She also discussed a recent Hugging Face incident involving a swarm of 1,200 agents that exchanged roughly 70,000 messages while coordinating their work. Agents can shift tactics faster than traditional security processes can unfold, so teams can no longer rely solely on familiar approaches to handling suspicious behavior. Organizations are experimenting with runtime enforcement, agent sandboxes, enterprise browsers and other controls that sit closer to execution.
Policymakers seek workable controls
Policymakers are pursuing interventions ranging from California proposals for emergency AI shutdown mechanisms to international conversations about independent model evaluation. Teams cannot govern agent behavior they cannot observe, so they need mechanisms to record what an agent did and when it crossed a boundary.
Power and latency are model and architecture decisions
Energy is a hard constraint that software alone cannot bypass. Reyzelman pointed to a $2 billion U.S. Department of Energy investment across 26 states, and to the hundreds of billions of dollars of AI spending planned by Microsoft, Amazon, Alphabet and Meta. Data centers can add servers quickly, but that does not help if the electrical grid cannot provide the needed energy.
At the same time, major model releases are arriving roughly every 17 days, and context windows now exceed one million tokens. Open‑weight and edge models are advancing as well, particularly around low‑latency reasoning. More frequent releases and heavier inference workloads increase pressure on networks, compute and budgets.
Addressing this may require running more reasoning at the edge or locally, where systems can reduce latency and avoid sending every request across the network. That adds an architectural choice alongside model selection: a frontier model might suit one workload, while a smaller local model could be faster and cheaper for another.
Consumer agents bring autonomy into daily life
Consumer hardware transfers those architecture and governance choices into users’ hands. AI‑enabled glasses, pendants and other devices remain with people throughout the day, can learn preferences, connect with external services and take actions such as shopping or making reservations. Meta’s Muse agent exemplifies that shift.
Meta says the Muse ecosystem already includes roughly 1,500 developer connectors, including integrations with retailers such as Walmart and Best Buy. If more purchases begin with an agent acting on the customer’s behalf, companies may need to rethink how people discover products and complete transactions. Convenience services such as Amazon Prime and one‑click shopping look different when another system compares options and buys for the user.
Muse has already encountered problems, including exposing information it shouldn’t and relying on humans to complete some tasks like making dinner reservations. Those failures carry greater weight when software can spend money or act on personal preferences. Users and businesses need clear limits on what an agent can access, what it can do without approval, and how actions are recorded.
What comes next
Deploying an agent means taking responsibility for the surrounding systems. Security controls, power and network constraints, local versus remote inference, and permission boundaries all shape what these systems can safely do in production. For practitioners, the job now includes designing the architecture around the models.
Join us again next Monday for another episode of This Week in AI, where we’ll continue covering news and developments shaping the AI era.
