Good morning. It's Monday, June 8, and we're covering how a Meta AI support agent was exploited, why small businesses are hiring teams of AI agents, and Anthropic’s call for a slowdown in AI development.

Plus: results from last week’s Real or AI challenge, and an original essay examining why model routing could become the dominant source of economic value in the AI era.

Anthropic argues that the biggest AI risk may arrive sooner than expected, warning that systems could begin improving themselves within two years. The company is urging leading AI labs and governments to consider a coordinated pause or slowdown in frontier model development while safety research and oversight catch up. Anthropic researchers cite the possibility of recursive self-improvement, though they acknowledge it remains theoretical. Critics question whether a global pause is feasible and suggest the proposal could also help preserve Anthropic’s competitive position.

Google disclosed in an SEC filing Friday that it will pay SpaceX $920M/month from October 2026 through June 2029 (~$30B total) for access to approximately 110,000 NVIDIA GPUs and related compute, framing the agreement as short-term bridge capacity to handle higher-than-expected Gemini Enterprise demand. The deal lands a week before SpaceX is expected to begin trading on the Nasdaq at a target ~$1.75T valuation in what would be the largest IPO ever, layering on top of last month's $1.25B/month Anthropic agreement and Google's separate $100B+ existing equity stake in SpaceX.

President Trump told reporters Friday he's discussing deals with AI companies that would let Americans benefit from AI's upside, following a CNBC report that his administration is exploring an equity stake in OpenAI to seed a "Public Wealth Fund" the company itself proposed, with proceeds potentially distributed directly to citizens. Sam Altman has reportedly been pitching the idea since early 2025, and the move aligns with Trump's existing pattern (a 10% government stake in Intel last year), while Senator Bernie Sanders is pushing a parallel proposal from the other side: a one-time 50% tax on OpenAI, Anthropic, and xAI payable in stock as all three target 2026 IPOs.

Researchers have used AI to create a vaccine antigen designed to protect against entire virus families, including future coronavirus threats that do not yet exist. Scientists at the University of Cambridge say the vaccine is the first to have its core antigen designed entirely by AI and then tested in humans. Early safety trials involved 39 participants, with a larger study of about 200 people underway to assess immune responses. The team is also applying the approach to influenza, bird flu, and Ebola-related viruses.

Anthropic warns AI may soon build its own successors, accelerating innovation while raising concerns about control, safety, and society’s readiness.

I’ve been speaking with a lot of Fortune 500 executives, and they are nervous. Annual budgets and multi-year contract commits are getting consumed in months, individual developers are spending thousands of dollars per day on coding agents, and public markets are beginning to question whether it’s acceptable to pay $100 for an internal powerpoint presentation. → Read the full article here.

Attackers reportedly used Meta’s AI-powered customer support agent to take over Instagram accounts by persuading the system to link profiles to attacker-controlled email addresses. The incident underscores a growing AI security challenge that extends beyond fears of highly advanced models such as Anthropic’s restricted Mythos system.

While recent debate has focused on AI capable of sophisticated cyberattacks, the Meta case shows that simple prompt-based manipulation can still produce real-world harm when AI agents are entrusted with sensitive tasks. As companies automate more customer support and operational workflows, securing AI systems against basic social engineering and authorization failures is becoming an increasingly urgent priority. → Read the full article here.

A growing number of small-business owners are using AI agent platforms such as OpenClaw to automate customer service, bookkeeping, marketing, scheduling, research, and other administrative work. The New York Times profiles entrepreneurs who have delegated significant portions of their businesses to AI “agents” powered by large language models from companies including OpenAI, Anthropic, and Google, with some reporting dramatic productivity gains and reduced staffing needs.

The technology remains unreliable and expensive at times, with users encountering unexpected failures, security risks, hallucinations, and prompt-injection vulnerabilities that can expose sensitive information. Despite those limitations, many early adopters view AI agents as a preview of how white-collar work may evolve, while economists warn that the same tools could accelerate job displacement, wage pressure, and inequality as automation expands beyond routine tasks. → Read the full article here. (Paywall)

A team led by Columbia's Dieter Egli posted a bioRxiv preprint June 1 reporting the first use of base editing on human embryos, making precise single-letter A-to-G changes in PCSK9 (cholesterol regulation) and HBG1/HBG2 (fetal hemoglobin, relevant to sickle cell and thalassemia). The work is being praised as more careful than the 2018 He Jiankui CRISPR-Cas9 attempt (He served three years in Chinese prison and recently told the NYT he stood by it), but the embryos showed mosaicism (some cells edited, others not), and Egli himself stresses the technology is not ready for the clinic because high doses of the delivery mRNA caused cells to stop dividing.

Bioethicists worry the preprint will spur a commercial rush even though Egli says that would be premature: Stanford's Hank Greely warns a millions-of-dollars IVF and genetic-testing lab could already attempt it, while UC Berkeley's Fyodor Urnov calls embryo editing for disease "a solution in search of a problem" because IVF plus genetic screening already prevent disease transmission, leaving the actual demand pointed at baby improvement. → Read the full article here.

Amazon Debuts Conversational Robot: New Proteus follows natural-language commands as Amazon expands automation and invests €10 billion.

Meta Uses Tents To Speed Data Centers: Meta is running data centers inside weatherproof tents rather than waiting years for permanent buildings.

SpaceX Faces S&P 500 Wait: S&P Dow Jones kept rules requiring a waiting period for newly public companies, delaying SpaceX's index inclusion.

Russia Targets Starlink Rival: Iks Holding says its Rassvet satellite internet service will begin testing soon and launch commercially in 2027.

GM Bets on New EV Battery: GM says its new battery development center will help cut EV costs nearly 10% and speed lower-cost battery launches.

Linux Foundation Tackles AI Costs: The new foundation aims to explain why enterprise AI spending is rising despite a 98% drop in token prices.

The image posted Wednesday was AI. Here’s how you voted:

Thanks for reading. See you next time!

— Matthew Berman, Nick Wentz & the Forward Future Team