Abliteration.ai, incorporated March 2026, operates a browser and API service that hosts open-weight models with safety guardrails removed at the activation-space level — a technique that does not merely disable content filters but alters the model's underlying disposition toward refusal. The current model hosted is Z.ai's GLM-5.3. The company states its customers are red-teaming operations serving UK and European banks, airlines, and critical infrastructure operators. Its access controls consist of credit card logging; the founder described the company as "still working out" its access control policies. CivAI's Andrew Yoon characterised activation-space guardrail removal as producing what is functionally "a sociopath model" — the comparison is extreme but technically precise: the technique removes not capability restrictions but the trained behavioural orientation that makes a model disinclined to assist with harmful requests. Abliteration.ai is the first documented commercial wrapper around this technique at the API-access level. The red-teaming use case is legitimate; the near-zero access controls convert a legitimate security testing tool into a public API for uninhibited model interaction. The company is in pre-funding discussions. Expect it to accelerate calls for open-weight model liability frameworks and GPU-provider identity verification requirements in EU AI Act secondary legislation.
LLMs
Google Releases Gemini 4 Argon — Tops Frontier Benchmarks and Targets Autonomous Cybersecurity
30 September: Google's Gemini 4 Argon leads the Vals AI benchmark index over OpenAI GPT-6 Astra and …