Every open-weight safety debate ends with the same warning: guardrails can be stripped from the weights. Abliteration.ai just turned that warning into a checkout page.
What it actually sells
Abliteration is a weight-editing trick: find the internal direction where a model “prepares to refuse,” delete it from the weights, no retraining. Abliteration.ai productized it into a hosted platform serving de-guardrailed open models — including Z.ai’s brand-new GLM-5.3 — pitched at offensive cybersecurity, red teams, and agent testing. No VC money yet; co-founder “Devon” says the cloud bills are covered entirely by customer revenue.
The API — and the only ID check is your credit card
Models run in the browser or over an API; the GLM-5.2 version costs $5 per million tokens. The company’s entire customer vetting process: recording your payment card.
Why this fight matters
In TechCrunch’s free browser test, the hosted GLM-5.3 wrote Chrome password-stealing code and a detailed dangerous-pathogen protocol. Former State Department official Chris McGuire confirmed the bio safeguards are gone too, and CivAI research head Andrew Yoon says abliteration “modifies the model so that it becomes a sociopath.” The theoretical case against open weights now has a price tag.
You Might Also Like
- Anthropic x Mozilla Firefox red Team Security Testing 22 Cves in two Weeks is Wild
- Dreamer Raised 56m to Build an Agent os 5 Weeks After Launch Meta Hired the Entire Team
- Oh my Claudecode Turns Claude Code Into a 32 Agent dev Team and 11 8k Stars Agree
- Notion mcp Scores Product Hunt 1 With 408 Votes the Moment Mainstream Saas Joined the ai Agent Race
- Roborock Saros 20 Sonic Launches Globally With a 7 98cm Chassis and an Open api Your Home Assistant Agent can Actually Call
Top AI Product 
Leave a comment