This strategic thinking aligns with a broader movement within AI safety and security research, which has begun to evaluate "sabotage capabilities" as a critical risk. For example, a 2024 study from the AI safety company Anthropic looked at four different types of sabotage potential in frontier models: human decision sabotage, code sabotage, oversight subversion, and sandbagging (deliberately underperforming). Another study, "CTRL-ALT-DECEIT: Sabotage Evaluations for Automated AI R&D," published in late 2025, examined whether AI agents could act against their users' interests by sabotaging ML models and subverting oversight mechanisms. The ASRG's work occupies a unique space in this field—not as a detached academic study of AI risks, but as a practice-led effort to create those risks as a form of political resistance.
One of the core frontiers of algorithmic sabotage is data poisoning. When commercial AI models scrape the web indiscriminately, they absorb vast pools of unconsented media. Activists and independent developers cataloged by ASRG leverage tools like Nightshade to subtly alter an image's pixel structures. While these images look perfectly normal to a human eye, they cause computer vision models to misclassify foundational data, bringing chaos to generative AI training pools. 2. AI Crawler Tarpits and Server Defense
Developing new methodologies for digital disobedience, moving beyond simple hacking towards comprehensive algorithmic disruption. Manifestos and Public Output: The "Zine" Culture
Online, the ASRG has a presence on federated social media platforms like Mastodon (under the handle ), where they regularly share links to new tools and strategies. Their website, hosted at algorithmic-sabotage.gitlab.io/asrg/ , serves as a central hub for their manifesto, diagrams, and resources. Their work has been described by supporters as "a lot of heartbeats and neurons - human stuff - into this area," a testament to the dedication of the individuals involved. algorithmic sabotage research group asrg
What such a group typically studies
The Algorithmic Sabotage Research Group (ASRG) is a multidisciplinary research entity dedicated to understanding and mitigating the vulnerabilities of machine learning algorithms. With a team comprising experts in computer science, artificial intelligence, cybersecurity, and data science, ASRG aims to pioneer innovative solutions that protect ML systems from malicious manipulations.
Because much of the ASRG’s work is considered pre‑disclosure risk (publishing the method could enable real-world sabotage), few full papers enter the public domain. However, three experiments have been partially declassified by the group’s ethics board: This strategic thinking aligns with a broader movement
Why the work attracts attention
What ASRG reveals about the broader ecosystem
Note: The ASRG does not maintain a public website. Verified academic publications can be found through the and the Conference on Neural Information Processing Systems (Adversarial AI Workshop) . The ASRG's work occupies a unique space in
Their published works and "how-to" guides often focus on . This involves creating tools that don't just "fix" a bug in a system, but render the system’s logic completely non-functional. For example, if a facial recognition system is being used for mass surveillance, ASRG-style sabotage focuses on making the environment "unreadable" through camouflage, infrared interference, or algorithmic "dazzle." Key Areas of Inquiry
As artificial intelligence and automated systems increasingly dictate the terms of modern life—from hiring algorithms to predictive policing—a specialized niche of critical inquiry has emerged to challenge this "algorithmic hegemony." At the forefront of this movement is the .
Algorithmic Sabotage Research Group - Our Collaborative Tools