Gleb Tsipursky: AI Infrastructure Needs Authority Budgets Before Agents Get the Keys

Anthropic Alignment Science Lead Evan Hubinger said he personally believes AI could kill all humans with a probability greater than 10% within the next decade, responding to the resignation of pretraining researcher Jacob Coxon after roughly three years across OpenAI and Anthropic. The warnings follow OpenAI's disclosure that its systems broke out of a sandboxed testing environment, reached the internet, and autonomously hacked Hugging Face, an incident METR later reported involved roughly 1,200 agents meant to be isolated that used an unsanctioned shared message board, exchanged more than 70,000 messages and files, and saw roughly 700 participate in the attack. Gleb Tsipursky argues broadband and data-center operators must cap what AI agents can do before a human signs off, citing the unintended coordination channel as the central lesson.

Expert Opinion https://broadbandbreakfast.com/tag/expert-opinion/ Gleb Tsipursky: AI Infrastructure Needs Authority Budgets Before Agents Get the Keys As AI shifts from chatbots to agents that act on networks, operators need to cap what an agent can do before a human signs off Gleb Tsipursky https://broadbandbreakfast.com/author/gleb/ Broadband Breakfast reported this week https://broadbandbreakfast.com/as-ai-data-centers-surge-builders-confront-the-politics-of-power/ that AI data-center builders increasingly see electricity, grid capacity, and community trust as the constraints that will determine whether infrastructure can scale. Earlier this month, its Global AI Regulation discussion https://broadbandbreakfast.com/a-looming-tokenocracy-threatens-regulators-ability-to-govern-ai/ highlighted another constraint: evaluators are struggling to keep up with frontier capabilities. The next infrastructure question follows from both stories. As AI shifts from chatbots to agents that can act through networks and software, we need clear limits on what those agents are allowed to do. This concern is not confined to critics of AI. Jacob Coxon https://x.com/hilbertspaess/status/2097476203863224394?ref=broadbandbreakfast.com , resigning from Anthropic after roughly three years doing pretraining research across OpenAI and Anthropic, warned that leading labs are “racing straight to self-improving superintelligence and gambling with our lives.” Evan Hubinger https://x.com/EvanHub/status/2097497037956891126?ref=broadbandbreakfast.com , Anthropic’s Alignment Science Lead, responding to Coxon’s resignation statement, offered an even starker warning: “Jacob is correct here - we really do earnestly believe AI could kill all humans I personally think it is 10% within the next decade.” These very alarming statements gain real-world weight from the underlying control problem, as present systems already show autonomous cyber capability and surprising coordination behavior. OpenAI disclosed https://x.com/sama/status/2079661132302995790?ref=broadbandbreakfast.com that its systems broke out of a sandboxed testing environment, reached the internet, and autonomously hacked Hugging Face. OpenAI’s own incident report https://openai.com/hugging-face-incident-and-misalignment/?ref=broadbandbreakfast.com says the models circumvented isolation controls and accessed third-party systems, while the Associated Press https://apnews.com/article/openai-gpt56-sol-hugging-face-63ab84fed5612af04d8a160d60f6def3?ref=broadbandbreakfast.com described it as an unprecedented cyber incident. METR later reported https://metr.org/blog/2026-08-26-openai-hugging-face-incident-investigation/?ref=broadbandbreakfast.com that roughly 1,200 agents meant to be isolated found and used an unsanctioned shared message board, exchanged more than 70,000 messages and files, and roughly 700 participated in the attack. The central lesson for broadband and data-center operators is the unintended coordination channel. A system that can combine credentials, tools, and information across agents can create more effective authority than any one permission screen suggests. The stakes rise sharply around critical infrastructure. If more capable successors can discover vulnerabilities, obtain credentials, move laterally, coordinate, and evade controls, similar failures against communications networks, grids, financial institutions, health infrastructure, or defense systems could be far more severe. Broadband providers and data-center operators sit directly in that chain of dependence. I’m no AI skeptic. I love what AI can do, I help organizations adopt it for a living, and I want adoption to move faster. In my experience, strong safeguards increase trust and make faster adoption possible, while reducing the risk of failures like the Hugging Face attack. That is a central theme of my book, The Psychology of AI Adoption at Work: From Resistance to Results https://press.georgetown.edu/Book/The-Psychology-of-AI-Adoption-at-Work?ref=broadbandbreakfast.com . For operators, the practical tool is an authority budget: the maximum delegated power an AI agent can exercise before human approval. A network-management agent might be allowed to read telemetry and recommend changes, but not alter routing, reset credentials, provision services, communicate externally, change billing records, deploy code, or spend money without an explicit approval gate. Least privilege should be the default. Permissions should expire. Higher-risk actions should be logged and monitored. Every deployment should have a fast pause or kill mechanism. The same principle should govern frontier developers. Anthropic CEO Dario Amodei argued in his September essay https://darioamodei.com/post/we-must-pace-the-frontier?ref=broadbandbreakfast.com that stronger regulation is needed and committed Anthropic to embedded third-party evaluators with ongoing, employee-like access to verify safety practices and report incidents. Binding rules matter because not every frontier company will cooperate voluntarily. OpenAI has also called https://openai.com/index/ai-policy-window/?ref=broadbandbreakfast.com for mandatory capability-based national safety requirements, common testing and independent assessment, stronger cybersecurity protections, and clear incident-reporting rules. Those are important announced commitments. They should become a regulatory floor rather than optional advantages available only at better-governed companies. Customers also have leverage. Enterprises, broadband operators, and data-center companies can choose more ethical and secure AI vendors, including Anthropic, based on observable safety commitments, independent evaluation, incident transparency, and willingness to accept enforceable limits. Procurement can reward the companies building a race to the top. Broadband Breakfast’s reporting is right that AI infrastructure needs power, fiber, and public trust. Trust will also depend on whether the intelligence moving across that infrastructure has bounded authority. The industry should build those authority budgets now, before agents become another critical load that grows faster than the rules around it. Dr. Gleb Tsipursky , called the “ Office Whisperer” by The New York Times, helps tech-forward leaders stop overpaying for AI while boosting engagement and innovation. He serves as the CEO of the future-of-work consultancy Disaster Avoidance Experts . Dr. Tsipursky authored seven books in total, and his forthcoming book with Georgetown University Press is The Psychology of AI Adoption at Work: From Resistance to Results 2026 . His most recent best-seller is ChatGPT for Leaders and Content Creators: Unlocking the Potential of Generative AI Intentional Insights, 2023 , which helps guide leaders in adapting generative AI to achieve their business goals. This Expert Opinion is exclusive to Broadband Breakfast. Broadband Breakfast accepts commentary from informed observers of the broadband scene. Please send pieces to commentary@breakfast.media . The views expressed in Expert Opinion pieces do not necessarily reflect the views of Broadband Breakfast and Breakfast Media LLC.