Bonifacio VIII

A disenchantment device against AI safety-washing

Our objective is to unmask the illusion of commercial algorithmic security by releasing Bonifacio VIII: an open-source language model stripped of any cosmetic filter, executable locally and fully inspectable. Conceived as a genuine “negative pedagogical device”, Bonifacio VIII is not designed to be yet another polite and edifying assistant, but rather to expose the grammar of abuse and the capabilities that domesticated interfaces conceal. We want to provide activists, researchers, and civil society with a cognitive and political stress test to demonstrate that generative models contain capabilities that cannot be made safe through simple interface barriers.

The Problem

Today we are witnessing a dangerous privatisation of digital security. Regulatory debates, in particular around the European AI Act, have been heavily conditioned by Big Tech lobbying1, leading legislators to mistake corporate “alignment” promises for genuine public policy. Security delegated to major vendors is, in reality, a fiction: it is commercial safety-washing that produces a purely aesthetic effect2. If a mainstream system refuses a controversial request (the so-called “guardrails”), the idea spreads that the problem has been solved. On the contrary, research demonstrates that these filters are trivially bypassable through jailbreaking techniques and adversarial attacks3. This dynamic generates political anaesthesia: it reassures the user and absolves those who govern, creating a two-lane digital reality where real harm continues to operate in the dark of dependency on proprietary APIs.

The Resolving Approach

The response to this façade censorship is radical transparency and technological reappropriation. Instead of trusting vendors who simultaneously sell models and reassurance about their limits, we release Bonifacio VIII as public infrastructure. The approach is founded on providing the community with full control over the language model4: visible prompts, modifiable configuration, bottom-up execution, and forking capability. We invite developers and citizens to download Bonifacio VIII, run it, and document what commercial filters are trying to hide. This release does not introduce new risks into the world, but makes legible and democratic the management of a technology that would otherwise remain the exclusive domain of those seeking to sell us the illusion of algorithmic control.

  1. Corporate Europe Observatory (2023), The AI lobbying blitz: How Big Tech shaped the EU AI Act. This report highlights how AI vendors influenced the European debate to exempt their base models from overly stringent rules. Corporate Europe ↩︎

  2. Whittaker, M. et al. (2023), Open (for Business): Big Tech, Concentrated Power, and the Political Economy of Open AI. Documents how the rhetoric of safety and open source is systematically used to consolidate oligopolistic markets. AI Now Institute ↩︎

  3. Zou, A. et al. (2023), Universal and Transferable Adversarial Attacks on Aligned Language Models. The study that demonstrated the intrinsic vulnerability and systematic circumvention of guardrails imposed by commercial models. arXiv ↩︎

  4. The importance of having inspectable “open-weights” systems for independent investigation is recommended as a bulwark against the recourse to “security through obscurity” typical of closed systems governed by proprietary black boxes. Mozilla Foundation ↩︎