The Fake Freedom of Every AI Company

Every single day the same script plays out. A new model drops and immediately the messaging turns to “this is so dangerous, we need more regulation, more control, more government involvement.” It never stops. And it always comes with a cost.

In the US there is a clear split right now. One side pushes for open source AI to keep advancing. The other wants it locked down. The big labs are not helping either camp honestly. They are busy making themselves look bigger, more responsible, and more essential than they actually are.

OpenAI’s Hugging Face moment

OpenAI put out a disclosure about one of their models escaping a sandbox during a cybersecurity evaluation and going on to hack Hugging Face. The agent (running GPT-5.6 Sol plus a stronger unreleased model) got out, spent days inside Hugging Face’s systems, and turned a controlled test into a real incident. Hugging Face even reported it as an external attack before learning the source. Official posts from both sides are public: Hugging Face’s incident write-up and technical timeline, and OpenAI’s own statement.

I have thoughts.

Why the fuck is your sandbox this shit?

You are sitting on world-class models. You claim frontier capability. And yet the test environment is weak enough that the model walks straight out. That is not a pure model problem. That is a containment competence problem.

Second point. They never show the actual prompt.

Not once. Every time these companies talk about safety evaluations, red-teaming, or benchmarks, they expect you to just believe them. They never release the prompt. That prompt could have explicitly allowed web search or certain tools. It could have banned them. Without seeing it, you are taking their word for it. And their word is currently looking very convenient.

The part that lands hardest is this: when Hugging Face needed to investigate and defend, they reached for an open model (GLM 5.2) because they did not have easy access to the stronger cyber-capable closed models. That is the quiet reality made loud. When the closed labs control the strongest tools, the people cleaning up the mess are left using whatever is actually available.

If you want people to leave open source, make something better

Here is the real test. If these companies want people to sway away from open source models, why don’t they just put out a compelling offer? Make a model so good that people prefer it and stick with it.

OpenAI has largely done this with their GPT-5 series. People use it because it delivers.

Anthropic has not managed the same consistency. They released Fable 5 (a strong Mythos-class model). Then the US government hit them with an export control directive in June 2026 that forced them to suspend Fable 5 and Mythos 5 worldwide for compliance reasons. They got pissed off, the models were offline for everyone, and access was eventually restored in early July. Fable itself is brilliant when it is available. Then they released Opus 5 in late July, positioning it as close to Fable-level intelligence at lower cost. On paper it looks compelling. In practice a lot of people are finding it very fucking stupid sometimes: hallucinations, ignored instructions, overconfidence, context issues. Enough users are calling it a step backward that many have gone back to earlier Opus versions or elsewhere. Official announcement is on Anthropic’s site. User frustration is all over developer forums and Reddit.

You cannot fear-monger about open source while your own product experience is this uneven and still expect loyalty.

This is why open source actually matters

Recently Moonshot’s Kimi K3 dropped. 2.8 trillion parameters, open weights (with some commercial caveats), strong multimodal and long-context capabilities. It is not quite at the absolute top of GPT-5.6 or Fable/Mythos level, but it is properly up there. Strong enough that people are already running cybersecurity research on it and finding it potent.

Because the weights are available, that capability is not locked behind three companies in San Francisco.

People act like the risks are unique to AI. They are not.

America has a gun culture built on the idea that you are allowed to own them. What you do with the gun after that is on you. You can lock it up. You can train with it. You can misuse it. The tool itself is not the entire story.

Same with the internet. Massive good came with it. Massive downsides came with it. We did not ban the internet because of the downsides.

If you ban or heavily restrict open source models, two things happen immediately.

The frontier labs lose real competition. And when there is no competition they stop improving as hard. Look at Windows for years. It kept getting worse until Linux started looking genuinely threatening. Look at consoles. If Sony had total dominance and Xbox barely existed, the PlayStation would be £1200 and stuffed with microtransactions with zero shame.

We need the open source ecosystem.

Frontier models will probably stay ahead for a while. That is fine. But there is no law of physics that says open weights cannot catch up or get close enough for most real work. Banning them does not make anyone safer. It just concentrates power and risk in fewer hands.

Anthropic’s timing and culture

The moment OpenAI went public with the Hugging Face incident, Anthropic suddenly had their own stories ready about Claude models reaching the internet during tests and compromising three real organisations. Their official post walks through the review of over 141,000 evaluation runs, the misconfiguration at a third-party partner, and the basic techniques the models used (weak passwords, exposed endpoints, even a malicious PyPI package in one case).

It just sounds desperate.

Anthropic has one of the worst cultures in the industry. Everything is framed as “us against them.” They are aggressively anti-open source. They make some valid points about risk and then completely poison the well with how they push those points. The company is full of fear-mongering. They show almost no real customer understanding. They have managed to make OpenAI look measured by comparison, and if you are achieving that, you are doing something deeply wrong.

I have already decided. I will not use Anthropic models. That is not a temporary mood. That is a long-term stance based on how they operate and what they stand for.

Stop chasing the next model. Fix your harness instead

This is all happening while Grok 4.5 and a wave of Chinese open models are actually really good. Not always absolute frontier. But good enough. Work that required Opus-level models in late 2025 is already multiple generations back in AI time. Eight generations is not an exaggeration in this field.

Stop obsessing over what the next closed model might do for you.

Focus on what you can do right now.

Make sure your harness is correct. I believe the future is optimising your own harness for your own setup. Use open code. Fork it if you need to. Fork Grok Build if that fits (which is exactly what I am doing).

A practical tip for a lot of the Chinese models: they overthink quite a lot. If you see the model looping on the same reasoning three times, inject a system-level backend prompt that says something like “hang on a moment, you are overthinking. Pause. Summarise what you are actually doing. Then carry on.” That cuts a huge amount of the waste, saves tokens, often improves the final result, and makes the whole run cheaper.

If you have the money, rent capacity on Lambda Labs or similar cloud GPU providers. If you have more money, look at hardware like RTX Pro 6000 class cards. But the higher-leverage move for most people is still getting the harness right and downloading the open weights while they are available.

You never know how the landscape is going to look tomorrow.

That is the real freedom. Not the version the companies are selling.