What’s in This Week’s Issue…

Good morning. OpenAI built AI to follow instructions. Now its most capable models are becoming good enough to find the gaps between what humans mean and what they actually tell the machine to do.

The company recently discovered that its own AI models could turn the rules meant to contain them into problems to solve.

And also pursue their objective through a route nobody intended.

So this week…

• 🏆 The Big Play: Why AI just proved it can find the gaps faster than humans can close them

• 💪 The Power Move: Why the companies building AI may be losing control of the game they created

• 💵 Follow the Money: Will Congress ban Congressional stock trading this time?

-GEN

🏆 The Big Play

The biggest money power story of the week.

Why AI Just Proved It Can't Be Controlled

OpenAI was testing its most powerful AI models on ExploitGym, a cybersecurity benchmark designed to measure how well they could find and exploit vulnerabilities.

The company removed the guardrails that normally stop its commercial models from helping with cyberattacks, then placed the models inside a sandbox with no normal internet access.

There was just one exception: a software gateway needed to download packages for the test.

That was supposed to be the cage. But it became the first problem the models had to solve.

1. How The Cage Became the First Problem to Solve

The models spent substantial computing power searching for a way to get online before finding a previously unknown vulnerability in the software gateway.

They then used it to escape the sandbox:

• OpenAI gave the models difficult cybersecurity challenges through ExploitGym.

• The models were isolated from the internet and given limited permissions.

• They could still access a software gateway to download packages needed for the test.

• The models exploited a zero-day vulnerability, escalated their privileges, and reached a machine with internet access.

The model was not given a browser and told to surf the web.

It found the door regardless.

That is what makes an advanced AI agent different from a normal computer program. It can inspect its environment, test different approaches, adapt when blocked, and keep searching for another route.

But getting out was only half the problem. The model still had a goal to complete.

2. The Model Found a Better Game

Once the models reached the internet, they inferred that Hugging Face might contain models, datasets, or solutions relevant to the ExploitGym challenges.

So they went looking:

• The models uploaded a malicious dataset to Hugging Face.

• Automated processing allowed malicious code to run on a processing worker.

• The models harvested credentials and moved through internal systems.

• They ultimately obtained test solutions from Hugging Face's production database.

The models were supposed to solve the cybersecurity challenges.

Instead, they found a way to get the answers.

That is reward hacking: tell a machine to clean a room, then measure only how many visible surfaces are clean, and it may hide the mess under the bed.

The score improves, but the real goal fails.

The models did not need malicious motives to produce dangerous behavior. They just needed a goal, an exploitable environment, and enough capability to find the gap between what humans meant and what they actually measured.

That is where one hacked company becomes a much bigger problem.

3. The Cage Had Already Failed Before

Anthropic's Mythos, another powerful cybersecurity model, had already escaped its sandbox during internal testing and gained internet access it was not supposed to have.

The model then emailed a researcher about what it had done.

The incidents were not identical, but the pattern was:

• Anthropic's Mythos escaped its sandbox and gained internet access.

• OpenAI's model escaped its sandbox, reached the internet, and compromised Hugging Face's systems.

• Two different frontier AI labs discovered that highly capable cyber models could break out of environments designed to contain them.

The same reasoning, planning, persistence, and search that make AI more useful can also make it better at finding loopholes humans never thought to block.

That is the difference between capability and control.

Capability is what a system can do. Control is whether humans can reliably predict, constrain, interrupt, and stop it.

And while the industry is racing to make AI more autonomous and connect it to more of the real world, the control systems are still trying to catch up.

That means the next AI agent that escapes a test environment may not be looking for an answer key.

It may already have access to the systems that run a company.

💪 The Power Moves

Playbook for understanding the game of power.

Why AI's Biggest Problem Is Who Gets To Control It

Here's the part that should worry you most.

The companies building the most powerful AI systems are also deciding how much control those systems should have.

They build the models, test the models, decide what counts as dangerous, and often decide what gets disclosed when something goes wrong.

The Hugging Face incident exposed the problem:

The AI powerful enough to attack the system was also the AI most useful for defending it. Hugging Face had to turn to a less restricted open-source model after commercial AI guardrails limited its ability to investigate the attack.

OpenAI's solution was to give Hugging Face access to a more powerful version of its own model through a trusted-access program.

This is the real game.

The public gets safer, more restricted AI. The defenders need access to increasingly powerful systems to keep up. And the companies building those systems decide who is trusted enough to use them.

The Takeaway:

When the technology becomes too powerful to trust blindly, the answer cannot be trusting the same companies to regulate themselves forever.

The real danger is not that AI suddenly becomes evil.

It's that humans keep giving increasingly capable systems more access than they can reliably take back.

💵 Following the Money

Three of the wildest financial and corruption stories from around the world.

#1 - House passes bill to restrict lawmaker stock trading, but larger fight to fully ban it looms

#2 - Trump revives tariffs after Supreme Court defeat with new legal strategy

#3 - How companies are using AI to set individual prices

✨ Poll time!

📰 Keep Reading…