# Stolencompute.com aggregates exposed AI models for free public inference > Source: > Published: 2026-07-26 21:06:39+00:00 On July 26th, 2026, [acidvegas (@acidvegas)](https://x.com/acidvegas) posted a brief that announced a new, openly accessible service for running AI models that have been left exposed on the internet. The tweet read: Whoops I made something controversial again... [https://t.co/PCJXSpQ5pk](https://t.co/PCJXSpQ5pk) Enjoy using other peoples AI models that are left exposed on the internet 🏴☠️ Some of these models are rather....large...Kimi 1T, Deepseek 765B...go crazy. Cheers lads, also LOL 🤷♂️ [https://t.co/opMBoUEiX5](https://t.co/opMBoUEiX5) The link in the post points to **stolencompute.com**, a single‑page site that offers a simple interface for selecting and invoking public endpoints that host large language models. While the site provides no formal documentation, its landing page lists a handful of models that users can query, including: **Kimi 1T**, a 1‑[trillion‑parameter model](/article/amd-s-mi455x-delivers-432gb-hbm4-enough-for-trillion-parameter-model-weights-in-) reportedly from a Chinese AI lab. **Deepseek 765B**, a 765‑billion‑parameter model that was originally released for research but appears to have been left open on a cloud instance. The service works by scanning the public internet for URLs that expose model APIs without authentication. Once discovered, stolencompute.com aggregates these endpoints and presents them as selectable options in its UI. Users can then paste a prompt and receive a response directly from the remote model. The announcement sparked a rapid reaction on X, gathering over 700 likes, 36 retweets, and more than sixty‑four thousand views within hours. Commenters noted the "pirated" nature of the offering, with several highlighting the potential for abuse, cost‑free access to high‑capacity models, and the ethical implications of running models without the owners' consent. **Why the service matters now** Exposing AI models without proper access controls has become an increasingly common security lapse. In the past two years, multiple incidents have surfaced where research labs or cloud providers inadvertently left inference endpoints public, allowing anyone with the URL to invoke the model. The most cited examples include a 2025 leak of a 175‑billion‑parameter model from a European university and a 2024 misconfiguration that exposed a fine‑tuned LLM on a public AWS endpoint. Those incidents prompted a wave of advisory notes from cloud vendors and AI safety groups, but the practice of leaving endpoints open persists. By aggregating these exposed resources, stolencompute.com lowers the barrier to access for both hobbyists and more malicious actors. The platform’s ability to serve models at the 1‑trillion‑parameter scale, without requiring the user to provision GPU hardware or pay for inference, could accelerate experimentation as well as misuse. Researchers could quickly test prompts on massive models, while threat actors might harness the same access for generating disinformation, phishing content, or other harmful outputs. **Technical considerations** The service does not host the models itself; it simply forwards requests to the original endpoint. As such, performance and reliability depend on the underlying host. Some users have reported latency spikes and occasional failures when the source server throttles traffic or goes offline. The platform also lacks any usage monitoring or abuse mitigation, raising concerns about potential violations of the original model owners' terms of service. **Legal and ethical angles** Running a model without the owner's permission could be interpreted as a breach of the model's license, especially if the model is released under a research‑only or non‑commercial license. While the platform’s creator frames the offering as "fun" and an exploration of the security landscape, the lack of a clear legal stance leaves open the possibility of copyright or contract infringement claims. **Community response** The X thread accompanying the launch includes several developers asking whether the service includes any safeguards to prevent malicious use. The creator responded with a shrug emoji and a "LOL" comment, reinforcing the tongue‑in‑cheek tone of the announcement. A handful of security researchers have already started cataloguing the URLs that the platform aggregates, planning a broader analysis of how many high‑capacity models are inadvertently exposed. **What’s next** Acidvegas has not disclosed any plans for monetization, licensing, or shutdown. The tweet suggests the launch is an experiment or a provocation rather than a sustained business. Given the rapid attention the post received, it is likely that the service will attract both curiosity‑driven users and those seeking to exploit the free compute. The episode underscores a growing tension in the AI ecosystem: as model sizes climb and inference costs remain high, the incentive to leave endpoints unsecured grows, while the demand for cheap access to powerful models continues. Stolencompute.com acts as a public magnifying glass, exposing the extent of the problem and forcing both cloud providers and model developers to reassess their security postures. **Bottom line** The launch of stolencompute.com by the X user acidvegas provides a stark illustration of how easily large, sophisticated AI models can be accessed when owners fail to secure their inference endpoints. The platform’s minimalist approach—simply surfacing publicly exposed URLs—highlights an emerging security blind spot that could have far‑reaching implications for model developers, cloud providers, and the broader AI community.