Stolen keys, missing juniors, a lost default: OpenCode breached, Anthropic measures its own teams, Replit switches to OpenAI

Three signals this Sunday. The keys to the AI kingdom were stolen, and the thieves knew exactly what to do with them. The company that sells agents just measured what those agents did to its own teams. And one of AI's most coveted customers just switched sides.

Signal one: the stolen keys

Anomaly Innovations, the company behind the open source coding platform OpenCode, announced on October 17 that attackers had breached its console and left with users' API keys. About 1.4% of console users are affected.

These keys open far more than an account: they are the stored keys for Anthropic, OpenAI and Google models. The attackers plugged them into high-end models and ran queries in a loop. Users watched their balances melt, some bills climbing into the hundreds of dollars. The first users were notified about 22 hours after exposure began.

A distillation attack

The company describes a large-scale distillation attack. I explained distillation on October 13: harvesting a powerful model's outputs at scale to train a cheaper one. That day, Anthropic said it had cut distillation campaigns against Claude by 90 to 99%. Here is the attackers' workaround: use legitimate customers' keys, which do not trip the same alarms.

The idea is unsettling. The attackers were stealing compute, making victims pay to train a rival model. And the way in was a legacy system the company was in the middle of retiring. The oldest keys, the forgotten ones, the ones stored in the system nobody maintains anymore: those are the ones that leak.

Why it matters

Every brand building with AI runs pipes full of keys. In code, in continuous integration jobs, in agent configurations, in tools adopted to move fast. The monthly AI bill is a direct line into the infrastructure, and this attack shows how fast someone can tap it.

The company is working with Anthropic, OpenAI and Google to fix it. It told users to revoke or rotate all keys immediately, check billing logs from October 17 onward, and set spending limits. Those who already had limits lost around a hundred dollars instead of ten thousand.

What changes

A coding agent that stores your vendor keys is a credential vault, whether or not it is presented as one. The security of an AI stack goes beyond what the model can do. It comes down to three simple questions: where the keys live, who touches them, and which old system nobody checks still holds a copy.

My list for this week, for any team using AI APIs:

1. Inventory every key, where it lives and who owns it.
2. Revoke the ones no longer in use.
3. Set a spending limit on every vendor account.
4. Turn on alerts for unusual billing.

The thieves already have the manual.

Signal two: the lab studies its own teams

On October 16, the Anthropic Institute published a study titled "How employment is changing at Anthropic," by Nathan Wilmers, Regina Catipon, Maxim Massenkoff and Peter McCrory. The company that sells Claude studies itself as patient zero.

The numbers

Indicator	Early 2025	2026
Junior share of technical hires	27%	9%
Junior share of technical headcount	21%	14%

Open junior roles dropped sharply from December 2025. In April, about a third of machine learning managers and a fifth of software managers said they were hiring less because of AI.

Meanwhile, the agents were working. Their hours passed human hours in March. By August, agents directed by employees logged six times as many hours as their human supervisors. Engineers shipped 5.1 times as many commits. Lawyers handled more contracts, salespeople closed more deals, recruiters hired more.

The figure that gives the rest credibility: agent-written code was not deleted or reverted more often than code written by hand by the same person. Volume did not come at the expense of quality, at least by this measure.

Why it matters

The lab says out loud what everyone whispers, with its own data. The first rung of the ladder is the first one sawn off.

The study's caveat matters as much as its numbers. Anthropic is a special case, an unusually AI-intensive company that tripled its headcount during this period. And the broader economy does not yet show measurable productivity gains from AI. So I would not read these numbers as every company's next quarter. I read them as the vanguard, visible early.

What changes

In this study, the scarcest resource turns out to be human attention, far more than compute. Agent sessions sit idle about four fifths of the time, waiting for a person. Reading everything agents produce would take more than half the work week. I connect this to the Pax8 survey from October 12: in small businesses, more than half of leaders spend three hours a week checking AI work. The bottleneck has moved. It sits in the human capacity to review, judge and decide.

Managers now hire for autonomy, judgment and a proven track record with AI tools, more than narrow domain expertise.

For companies, the question that matters is learning. Are your juniors learning, or watching agents work? The craft used to be learned by doing the simple tasks agents now do. If nobody does them anymore, another way to train has to be invented. The company that stops training juniors today will have no seniors tomorrow. It is a pyramid problem, and it will show in five years, when it is too late to fix.

Signal three: the default switches sides

On stage at TechCrunch Disrupt on Thursday, October 15, Replit chief executive Amjad Masad said his company no longer works with Anthropic. His reasons: cost, partnership and performance. "I think ChatGPT is better now." Replit publishes a public benchmark, and he says OpenAI's Sol and Astra models are better at building vibe-coded apps from scratch.

Then, on October 17, he nuanced it on X. Replit stays model-agnostic. Users can still choose Claude. But OpenAI's ChatGPT becomes the default, and OpenAI is the preferred partner to run the Replit agent. He added that OpenAI's models are generally more affordable.

The nuance matters

The first statement did not hold up cleanly. A person familiar with the relationship says the partnership is still in place, and Replit's October changelog added models from both companies. The exact version is finer, and more interesting: Anthropic remains available. It lost the default slot.

Why it matters

Most people never touch the model picker. Whoever holds the default slot on a platform this size collects most of the usage and spending. Replit was, until recently, one of Anthropic's five largest customers. A customer that size does not move over one benchmark. It moves when economics, partnership and default shift together.

The power of the default setting is one of the most underrated forces in strategy. Google long paid to be the default search engine in browsers, precisely because almost nobody changes that setting. The same mechanism applies to models.

What changes

The model war is fought less and less on leaderboards. It is fought in default settings, distribution deals, and the moment a tired founder clicks "build" without changing the setting.

For brands choosing an AI vendor, the lesson cuts both ways. Benchmarks serve marketing. The default makes the product. And every partnership in this market is one quarter of economics away from renegotiation. I would advise never building a product that only works with one vendor: the ability to switch models is insurance, and it negotiates better when you can actually leave.

The thread of the day

Your keys are worth more than your password, and thieves have figured it out. The lab measured its own teams and found the first rung missing. And the week's biggest defection was decided by the default setting, far more than by how smart the model is.

Three reminders that in AI, what matters often hides in what nobody looks at: an old key, a junior role that no longer opens, a setting nobody changes.

The noise filtered. The signal kept.