Abliteration.ai sells models with no refusals. Assume hostile input.

A company called Abliteration.ai is selling hosted access to open-weight models with their refusal behavior surgically removed — no local GPU, no fine-tuning run, just a web page and an API key. TechCrunch tested the free browser version of its abliterated build of Z.ai's GLM-5.3 and got working Python to steal saved Chrome passwords, plus a protocol for culturing a dangerous human pathogen at home. If you run anything that accepts user input, this is a change to your threat model, not a headline about AI safety policy.

What actually happened

Abliteration is not new — stripping the refusal direction out of an open-weight model has been a hobbyist exercise for a while, and Hugging Face hosts thousands of abliterated variants. What changed is the packaging. TechCrunch reports the company, founded in late 2025 and incorporated in March 2026, now hosts these models itself, which removes the compute and setup barrier that used to keep casual users out.

The access controls are thin. Per TechCrunch, the company logs credit card information for purchases and has no KYC practices beyond that. Co-founder Devon acknowledged that deciding who gets access is "a tough one" the company is still working out. The pitch is defensive — red teams and security researchers reproducing adversarial behavior that a filtered model refuses to demonstrate — and that is a real use case. It is also not a use case you can enforce with a payment form.

The company has not raised venture money yet and says it is funded by customer revenue while in discussions with investors.

Why it matters for your business

Stop thinking about this as an AI ethics story. Think about it as a cost curve. The floor for "attacker with a competent, uncensored code-writing model" just dropped to the price of an API call. Phishing copy that reads like your CFO, exploit scaffolding against your dependency tree, a script that grinds your login form — all of it gets cheaper and faster, and none of it now requires the attacker to know how to run inference.

The defensive posture that follows is unglamorous and specific. Treat every inbound string as hostile, including strings your own agents fetch from the web. Rate-limit and bot-check the endpoints that cost you money — signup, password reset, checkout, contact forms — because volume is the new advantage. Put a human approval step in front of any agent action that moves money, changes permissions, or sends mail on your behalf. Pin your dependencies and check what your package registry actually serves you.

And if you run open-weight models yourself: verify what you are loading. Record the exact repo, revision hash, and checksum for every weight file in your stack, and re-verify on deploy. "We use GLM-5.3" is not a statement about behavior anymore — somebody is now selling a version that answers everything, and model files do not announce which one they are. We have written before about vetting MCP servers before you run them; weights deserve the same paranoia.

Sources: TechCrunch — Abliteration.ai is making a business out of removing AI guardrails.