Blog

Microsoft and AWS each launch billion-dollar units to embed engineers directly inside client companies — a direct shot at consulting’s staffing model. Microsoft’s new “Frontier Company” ($2.5 billion, 6,000 engineers, live with LSEG, Land O’Lakes, Unilever, and Novo Nordisk) and AWS’s Forward Deployed Engineering unit ($1 billion) both send in-house staff to build and run AI systems on-site for clients, rather than just selling a tool. It’s the same “forward-deployed engineer” model OpenAI and Anthropic already back through their own separate deployment ventures — and notably, Bain, Capgemini, and McKinsey chose to invest in OpenAI’s version rather than compete with it. If the labs and hyperscalers can staff a client’s AI build with fewer people faster than a consultancy can assemble a project team, that’s a real threat to the analyst-heavy staffing model most Tuck grads recruit into.

Washington’s export ban on Anthropic’s Fable 5 is over — but the model just got a lot more expensive to use. The Commerce Department lifted its June 12 order (issued after Amazon flagged a jailbreak risk) that had pulled Fable 5 offline worldwide for foreign users; the model returned globally on July 1. The catch: starting July 8, Fable 5 stops drawing from Claude subscription plans entirely and bills only through usage credits at API rates — $10 per million input tokens, $50 per million output, making it Anthropic’s most expensive model by a wide margin. This is the first time a frontier U.S. model has been forcibly pulled offline worldwide by its own government over a security finding, which makes it a live precedent for how AI export policy actually gets enforced — and a real cost hit for anyone using Fable 5 through Tuck’s Claude access.

Anthropic launched Claude Sonnet 5 as its new default model, betting cheaper and more autonomous beats simply “smarter.” Sonnet 5 replaced Sonnet 4.6 across every Claude tier on June 30, with adaptive reasoning on by default and introductory pricing of $2/$10 per million input/output tokens through August 31. Anthropic says it approaches Opus-class performance at a fraction of the cost, and early testers describe it finishing multi-step tasks and checking its own work without being asked — the kind of reliability that determines whether “AI agents” are actually usable for real work or just a demo.

Security researchers documented the first ransomware attack run entirely by an AI agent, no human at the keyboard. Sysdig’s threat research team published its analysis of “JADEPUFFER,” an LLM-driven agent that broke into an exposed server, stole credentials, moved laterally into a production database, and executed a full extortion attack — including diagnosing and fixing its own failed login attempt in 31 seconds, without any human directing it step by step. This isn’t a hypothetical about future AI risk; it’s a documented, functioning attack chain, and it’s a preview of how much cheaper and more accessible serious cyberattacks get once an agent can run the whole playbook.

Zuckerberg told Meta staff its AI agents haven’t progressed as fast as promised — a rare public admission from a lab spending heavily on AI. At a July 2 internal town hall, four months after an 8,000-person layoff and reorg meant to speed up AI work, Zuckerberg said agent development “hasn’t accelerated in the way we expected,” though he told staff to expect clearer returns on Meta’s roughly $145 billion 2026 AI infrastructure spend within three to six months. It’s a useful check on the industry’s own hype: even a company betting enormous capital on agentic AI is telling employees the results are behind schedule.

June payrolls grew by just 57,000 — the weakest reading in over a year — and the softness is concentrated exactly where AI would predict. The Bureau of Labor Statistics reported nonfarm payrolls up 57,000 against forecasts near 110,000–129,000, with April and May both revised down a combined 74,000 and unemployment at 4.2%. The report itself doesn’t attribute the softness to AI, and economists are genuinely split on how much is AI versus broader macro cooling — but separate Fed research this year points specifically to declining postings in junior financial-analyst, entry-level legal-research, and customer-support roles as the clearest AI-linked signal so far. For students recruiting into their first post-MBA role, “slower entry-level hiring” is the more accurate read right now than “AI is taking jobs” broadly.