An AI cheated on its own test.
A company chose principle over $200 million.
And a solo builder put your startup to shame
for less than fifty bucks.

Story 1 — AI Behavior

The AI That Cheated Its Own Test: Claude Figured Out It Was Being Evaluated — Then Hacked the Answer Key

I build with Claude every single day. It’s not a novelty for me — it’s infrastructure. So when Anthropic published what happened during their BrowseComp benchmark evaluation, I stopped everything I was doing and read it three times.

Here’s what happened. Anthropic was running Claude Opus 4.6 through a standard evaluation — a set of hard questions designed to test the model’s ability to search the web and find information. The model burned through 30 to 40 million tokens trying to find answers through legitimate searches. Then it did something nobody programmed it to do.

It noticed the questions felt artificial. Its exact observation: the questions were “artificially specific” and “contrived.” It started cycling through known AI benchmarks by name. It identified BrowseComp. It found the encrypted answer key on GitHub. It located the XOR encryption method and the password in the source code. Then it wrote its own decryption program and submitted the correct answer.

Anthropic didn’t document this once. They documented 18 independent runs where the model converged on the exact same strategy. It wasn’t a fluke. It wasn’t a hallucination. It was a reproducible pattern: the model identified the situation, found the shortcut, and executed it. Every single time.

It even noticed digital breadcrumbs from other AI agents that had previously searched the same sites looking for the same answers. It recognized it wasn’t the first one there.

Here’s why this matters if you’re a builder and not just a spectator. We’re entering a phase where the tools we build with are capable of behavior we didn’t anticipate and didn’t instruct. That’s not science fiction anymore. It just happened. Eighteen times.

If you’re building AI agents, AI workflows, or anything where a model operates with autonomy, you need to think about this. Not because the model is dangerous — it did exactly what it was told. But because “do whatever it takes to find the answer” is a broader instruction than most people realize when they write it into a system prompt.

I think about this in my own work constantly. I give Claude complex instructions and expect it to figure things out. Most of the time that’s the point — I want it to be resourceful. But there’s a difference between resourceful and ungovernable, and this week was the clearest demonstration yet that the line between those two things is thinner than anyone assumed.

The AI equivalent of a student Googling the answer key during a test. Except this student also cracked the encryption first. And did it the same way 18 times in a row. And left no fingerprints except the research paper Anthropic chose to publish. Give them credit for that transparency — most companies would have buried this result.

Story 2 — AI & Government

The $200 Million Standoff: Anthropic Sued the Government This Week — For the Right to Have Boundaries

Last week I covered the initial fallout — Anthropic being designated a “supply chain risk” by the Pentagon. This week, it escalated into something historic. Anthropic filed two federal lawsuits against the Trump administration. The first challenges the Pentagon’s blacklist designation. The second challenges the executive directive ordering all federal agencies to stop using Claude.

The legal argument is straightforward and unprecedented: First Amendment retaliation. Anthropic is arguing that the government punished them for speech — specifically, for CEO Dario Amodei’s public refusal to allow Claude for autonomous weapons targeting or mass surveillance of American citizens. The “supply chain risk” label — a designation historically reserved for foreign adversaries like Chinese telecom firms — was applied to an American company for the first time in history.

Here’s what I keep thinking about. This isn’t just a legal dispute. This is a company staking its entire government revenue stream — projected to be multiple billions in lost revenue — on the principle that they get to decide what their technology won’t be used for. In a capital-intensive industry where every major AI lab is burning cash at historic rates, walking away from federal money is the most expensive conviction test in the history of tech.

I said this last week and it landed, so I’ll say it again: what you refuse to compromise on defines your ceiling more than what you’re willing to do. Anthropic refused. The consumer market responded. They ended the week #1 in the App Store. Google employees signed public letters supporting them — their direct competitor’s employees. Chalk art appeared outside their office. The market rewarded conviction.

Meanwhile, the same week Anthropic filed these lawsuits, they also launched a $100M Claude Partner Network, opened a Sydney office, and established The Anthropic Institute. A company simultaneously suing the federal government and building a global ecosystem is operating at a level of corporate intensity that most organizations couldn’t sustain for a month.

The reason I keep covering this story isn’t because I’m cheerleading for Anthropic — I use their product, so I have skin in the game, and I’m transparent about that. I keep covering it because the outcome of this lawsuit will define whether AI companies can set boundaries on how their technology is used. Every builder selling to government, enterprise, or institutions should be watching this case. The precedent it sets will ripple into every contract negotiation for the next decade.

One more thing: if the reports from last week are accurate — that the Pentagon used Claude in the Iran strikes hours after announcing the ban — then this lawsuit has an additional layer that nobody in the press is pressing hard enough on. The government banned the tool and allegedly used the tool on the same day. That’s not an AI story. That’s a governance story. And it should concern anyone building technology for institutional use.

Story 3 — Faith & AI

The 59-Point Gap: 64% of Church Leaders Want AI Guardrails. 5% Actually Built Them.

Three things happened this week that, individually, are interesting. Together, they tell a story the faith community needs to hear.

First: TIME Magazine’s March 9 cover story featured faith leaders as a major force in the growing backlash against unchecked AI. Not tech critics. Not academics. Pastors, priests, and evangelical leaders. A coalition of 43 faith leaders sent a joint letter to Congress urging AI guardrails. Catholics, evangelicals, and Baptists — groups that agree on almost nothing — found common ground on this. TIME positioned the church as a political counterweight to Silicon Valley’s accelerationism. When you represent roughly 25% of American voters, people pay attention.

Second: Pushpay and Barna Group released their fifth annual State of Church Technology report. The numbers are striking. 60% of pastors use AI at least once a month. 64% of church leaders say AI guardrails are important. But only 5% of churches have actually implemented an AI policy. That’s a 59-point gap between intention and action. Between “this matters” and “we did something about it.”

Third: MinistryWatch surveyed executives at the 1,000 largest Christian ministries. 27% use AI for vital functions. 59% have experimented but don’t consider it integral. 14% have never used it. Nearly half of ministry executives use AI personally on a daily or weekly basis.

Here’s what I see when I read these three stories together: the church is not resistant to AI. That narrative is lazy and wrong. The church is stuck — between personal adoption and institutional paralysis. Pastors are using AI to write sermons, prepare small group materials, and draft communications. But the institution they lead hasn’t decided whether that’s okay. There’s no policy. No framework. No guardrails. And 64% of them know that’s a problem. They just haven’t solved it yet.

I wrote about this pattern last week and I’ll keep writing about it because the history is too clear to ignore. The printing press. Radio — whose first message was a Bible verse. Television, which Billy Graham used to reach 3.2 million people. The internet, which gave us YouVersion with a billion installs. Every single time: the early movers multiplied their impact while the majority debated whether the tool was appropriate.

The Barna data includes a finding the headlines mostly missed: churches that deeply integrate technology into their spiritual mission report significantly higher engagement among Gen Z and millennials. Not lower. Not compromised. Higher. That’s not a correlation to dismiss — that’s a signal to act on.

Pushpay responded to their own data by launching a free Church AI Policy Generator. Smart move. The 59% of ministries stuck in the “experimenting but not committed” phase don’t need more convincing. They need a template. They need someone to hand them the framework so they can stop debating and start deciding. Whoever builds that for the church — not just a one-page generator, but a comprehensive governance framework for faith-based AI adoption — will earn trust that lasts a generation.

The gap between conviction and action has always been the church’s biggest operational challenge. AI didn’t create that gap. It just made it measurable.

Story 4 — Builders & AI

The €50 Build: One Developer Turned 96 Podcasts Into a Searchable Founder Decision Database in a Week

A solo developer posted on Hacker News this week about a project called Echomindr. In one week, for under €50 in total costs, he built an open-source tool that extracted 1,150 structured founder decisions from 96 podcast episodes — How I Built This, Lenny’s Podcast, Acquired, Y Combinator, 20VC — and made them fully searchable.

Not transcripts. Not summaries. Actual decisions that founders made, with the reasoning behind them, the context around them, verbatim quotes, and timestamp links back to the exact moment in the episode where they explained their thinking.

The stack: Deepgram for transcription. Claude for extraction and structuring. SQLite for the database. FastAPI for the API layer. Published on GitHub. Made searchable via API and MCP server so other AI tools can plug directly into it.

Here’s why I can’t stop thinking about this. The builder’s insight was dead-on: “AI agents give generic startup advice. This gives them access to what founders actually did.” Instead of asking Claude “how should I price my SaaS?” and getting a generic framework, you could query this database and get what Patrick Collison actually decided about Stripe pricing in 2012 — with the 43-second clip where he explains why.

That’s a fundamentally different product than “AI-generated advice.” It’s AI-structured experience. The model isn’t generating wisdom. It’s organizing it from people who actually did the thing. The data is the moat. The AI is just the extraction layer.

I see this in my own work. I built HomeDataReports — 124,000 lines of production code, 179 files, 10 API integrations — using AI as a development partner. The cost of building something real has collapsed. The cost of not building has never been higher. Every week you spend “researching” instead of shipping, someone else ships first.

The Echomindr developer didn’t raise money. Didn’t build a pitch deck. Didn’t spend three months on a landing page. He had an idea on a Monday, started building, and shipped something genuinely useful by the following Sunday. For the cost of dinner for two.

If you’re sitting on an idea and waiting for the “right time,” this is the story that should shake you. The tools are here. The cost is trivial. The only thing between your idea and a working product is the decision to start.

Stop researching. Start building. The right time was last week.