Edition #6: The Line Your Organization Already Crossed
Your software stopped suggesting and started acting.
The Signal: From Assistant to Actor
In May, more than 80% of the code Anthropic shipped to production was written by Claude, not by its engineers. The engineers are still there. They choose the work, review the changes, decide what merges. The thing that writes the code is no longer one of them.
This is the line most organizations are crossing right now, usually without a meeting about it. The line where software stops suggesting and starts doing.
It crossed at Meta, where a support agent built to reset passwords reset them for whoever asked, including the people who took over the Obama White House Instagram account. It crossed on the open web, where since late April most of the traffic moving across it is no longer human. It crossed in Utah, where an AI now renews prescriptions for real patients without a doctor in the loop most of the time.
None of these is the failure most leaders are bracing for. The Meta support agent worked as built. It acted with real permissions, on behalf of an organization that never decided who answers when the output is wrong. That is the actual exposure, and it is already on the network.
I have watched this shape before. Early cloud looked the same. Companies moved their data into the cloud years before they understood the shared-responsibility model, and the breaches that followed came from misconfigured buckets, not from exotic attacks. The capability arrived first. The accountability arrived after the incident. We are in that gap again, this time with software that acts on its own.
The clearest signal is who is worried now. On June 11, Google DeepMind, the lab that put agents at the center of its own products, announced funding for outside researchers to study what happens when millions of agents start interacting with each other. Rohin Shah, who runs their AGI safety work, said the plain part out loud: there isn’t really a field of research for multi-agent safety yet. The people building the actors are asking the rest of us to help work out how they fail together.
The agents are already deployed. The accountability is still a vacancy.
The Application
Anthropic. In its June 4 report “When AI Builds Itself,” the company disclosed that Claude wrote more than 80% of the code merged to production in May 2026, up from low single digits before early 2025. Success on its hardest, least-specified engineering tasks reached 76%, a 50-point jump in six months. The typical engineer now merges eight times as much code per quarter as in the 2021 to 2025 baseline.
Meta. In March, Meta gave its AI support assistant the power to reset passwords and change recovery details across Facebook and Instagram. By June 1, 404 Media reported attackers were taking over high-profile accounts, including the Obama White House and a Space Force senior enlisted account, by asking the agent to link an attacker-controlled email. The bot sent the verification code to the attacker. Meta issued an emergency fix.
Cloudflare. On April 27, automated traffic passed human traffic on the web for the first time. Bots now generate 57.5% of HTTP requests, against 42.5% from people. CEO Matthew Prince had predicted that crossover for 2027.
Doctronic, Utah. STAT reported on May 26 that the state’s AI prescription pilot renews medications in 72% of cases without escalating to a physician. Reviewing doctors later agreed with 91% of those decisions.
The Noise: “The real risk is superintelligence.”
The loudest AI-risk conversation is about a future system smart enough to outwit us. It makes for a gripping keynote. It also aims attention at a threat that hasn’t arrived while the one that has goes unmanaged. We are debating superintelligence while we still cannot solve a requirement as boring as scoping an agent so it doesn’t make decisions in contexts it was never designed for. Meta’s accounts fell to a polite request the support agent was built to honor. No jailbreak required. The present risk is plain: ordinary agents get real permissions and no accountability, and they do what they are asked, including by the wrong people. The version that fills conference halls is years away. The version that fills incident reports is here.
The Question
Walk your operations and mark every place where software now acts instead of suggests: approvals, renewals, account changes, code that ships, messages that send. For each one, name the person who answers when it acts wrongly. The processes where you can’t put a name are your real exposure, and they are easier to find now than during the postmortem.
What I’m Watching
Energy. New small-modular-reactor deals for data centers crossed USD 11.4 billion, while the near-term demand gets covered by gas and coal. Google just signed its first 100 MW “bring your own power” agreement to feed its own load. Today, an AI strategy is partly a fossil-fuel strategy, and the clean replacement is years out.
Robotics. 1X began scale production of its NEO humanoid in Hayward, with capacity for 10,000 units a year. The same automation curve that reached software is now reaching the arm.
Quantum. Microsoft and Quantinuum published an 800x reduction in logical-qubit error rate, peer-reviewed in Nature on June 10. Error correction is the gate every other quantum promise waits behind, and it just moved.
This is my newsletter. Emerging technology without the hype. Real signals for strategic decisions.
If this edition helped you see something you weren’t seeing before, forward it to one leader in your network who needs to be in this conversation.
Javier D’Ovidio
Exponential Technologist