Abliteration.ai has turned a once do-it-yourself model-modification technique into a hosted commercial product. According to The Decoder, the U.S. startup sells access to modified open-weight models whose trained refusal mechanisms have been stripped or suppressed, rather than asking customers to download weights and run their own infrastructure. The current product is “abliterated-model-large-v2,” which The Decoder reports launched in late August and is a modified version of Z.AI’s GLM-5.3. The model is designed to refuse sensitive requests far less often than the base model. The company markets that behavior for offensive cybersecurity, AI red teaming, agent testing, and trust-and-safety work, including vulnerability reproduction, proof-of-concept exploits, malware analysis, and simulated phishing. The underlying technique is called abliteration. As The Decoder describes it, the process identifies internal activation patterns associated with refusal behavior, then changes the model weights to suppress those patterns. That makes it different from a prompt jailbreak: the refusal behavior is altered in the model itself, not merely bypassed in a single interaction. The commercial packaging is the important shift. The Decoder notes that developers have published abliterated models on Hugging Face for years, but Abliteration.ai does not offer its modified weights as a public download. Instead, it hosts the model and sells access as a service. The standard price for the abliterated GLM-5.3 model is $5 per million input or output tokens, according to the report. Abliteration.ai says the modified model keeps most of its coding, cyber, and agentic capability. The Decoder cites the company’s own evaluations showing 84.5% on CyberGym, 41.8% on Terminal-Bench 4.0, and 105 solved ExploitGym tasks in two hours for the abliterated GLM-5.3 version. The same report notes that the model does not lead every benchmark in Abliteration.ai’s table: GPT-5.5 is listed higher on CyberGym at 85.6%, while GPT-5.6 Sol and Fable 5 score well above it on ExploitGym. Those comparisons come with a caveat. The Decoder reports that Abliteration.ai itself acknowledges the benchmark scores use different harnesses and compute budgets, limiting direct comparability. For operators, that means the benchmark table is useful as a directional claim about retained capability, not as a clean ranking of model performance. The legal path runs through open-weight licensing. The Decoder reports that GLM-5.3 combines coding, agentic, and cyber performance with open weights and a commercially usable license, and that Z.AI permits modifications, derivatives, and commercial “Model as a Service” offerings. That license structure lets Abliteration.ai modify GLM-5.3, host the resulting model, and sell access to it. The security trade-off is explicit. The same properties that make a low-refusal model useful for red teams and agent-safety testing also lower the barrier for problematic use. The Decoder’s summary says journalists were able to generate malware instructions using the service, underscoring the risk that hosted access can make previously niche model modification workflows easier to operationalize. Who benefits: Security teams, red-team vendors, and organizations testing AI agents may benefit from easier access to models that refuse fewer cybersecurity prompts. Abliteration.ai also benefits from turning a technical workflow into a hosted service priced per token. Who's exposed: Defenders face a lower barrier for actors seeking models that are less likely to refuse harmful cybersecurity requests. Organizations deploying AI agents are part of the risk picture because the service is marketed for agent testing as well as red-team and trust-and-safety work.