A central finding of the ASRG is that any algorithmic system resistant to red-team auditing is vulnerable to blue-team sabotage. We formalize this as the Dual-Use Audit Corollary:
For any deployed classifier ( C ) with a rejection threshold ( \tau ), if there exists no adversarial perturbation ( \delta ) such that ( C(x+\delta) ) falls into a human-review bucket, then ( C ) is either a constant function or has been overfitted to the point of practical uselessness.
In practice, the ASRG has demonstrated that injecting carefully crafted "grey noise" (e.g., adding 0.0001% Gaussian noise to an insurance application’s timestamp) can shift a denial into a "manual review required" state. This is not breaking the system; it is revealing the brittleness of its confidence intervals.
Most red-teaming exercises test how an algorithm handles malicious inputs. The ASRG flips the script: they test how an algorithm handles malicious internal states. Their red teams play the role of a rogue developer or compromised data source. They ask: If I wanted this AI to fail in six months, how would I subtly corrupt the retraining pipeline today? This proactive research has produced a library of over 200 "sabotage patterns," from gradient poisoning to delayed-action trigger conditions.
The ASRG operates under a strict Hippocratic Oath for Sabotage:
Our goal is friction, not fracture. We aim to lower the velocity of automated injustice until the human-in-the-loop can catch up.
No discussion of the Algorithmic Sabotage Research Group is complete without addressing the controversy it courts. Critics raise three primary objections:
1. The Blueprint Problem. By publishing detailed analyses of sabotage techniques (even for educational purposes), the ASRG arguably provides a playbook for malicious actors. A competitor could read an ASRG white paper and replicate the "geo-loop" sabotage in their own system. The ASRG counters that sunlight is the best disinfectant; without public disclosure, regulators would never know what to look for.
2. Vigilante Forensics. The ASRG is not a law enforcement body. Yet, its reports have been used in shareholder lawsuits and regulatory hearings. Critics argue that the group’s lack of formal legal process (e.g., chain of custody for data) could lead to false accusations. The ASRG maintains a strict policy of "attribution without accusation"—they identify the presence of sabotage mechanisms but refuse to name specific corporate actors unless the pattern is independently verified by a government agency.
3. The Saboteur’s Dilemma. Perhaps the most biting critique is that ASRG members themselves possess the exact skills needed to commit algorithmic sabotage. A former member of the ASRG’s red team was banned in 2024 for selling a zero-day sabotage exploit on the dark web. The group acknowledges this risk and has since implemented psychological screening and blind-review protocols, but the shadow of the "reformed hacker" remains.
The group theorizes the legality and ethics of sabotage. They argue that sabotaging an algorithm is a form of civil disobedience, particularly when the algorithm itself is deemed unjust (e.g., a biased predictive policing tool).
The Algorithmic Sabotage Research Group (ASRG) is an artistic research collective and theoretical platform dedicated to investigating the politics of algorithms. Rooted in the traditions of tactical media, critical theory, and digital art, the group explores how "sabotage" can be used as a methodology to disrupt, expose, and challenge the power structures embedded within contemporary computational systems.
The ASRG is not a traditional scientific laboratory; rather, it functions as a hub for interdisciplinary inquiry, bringing together artists, hackers, writers, and theorists to examine how code influences society, labor, and human behavior.
The ASRG engages in "Speculative Fabulation" and practical experiments. Some notable areas of focus include:
