| Attack Surface | Target | ASRG Research Focus | |----------------|--------|----------------------| | | Labeling services (e.g., Mechanical Turk) | Subversion of annotators : paying workers to systematically mislabel a specific class (e.g., all "pedestrian" as "street sign"). | | Model Registry | Hugging Face, internal model stores | Trojan model uploads : publishing a "helpful" fine-tuned model that contains a logic bomb. | | Inference API | Public-facing ML endpoints (GPT, Claude, Gemini) | Extraction via sabotage : crafting queries that force the model into a repetitive, resource-exhaustive loop (a new form of algorithmic DoS). | | Continuous Learning Pipeline | Online retail, fraud detection | Drift injection : feeding a slow, plausible shift in input distribution so the model gradually becomes racist, sexist, or financially reckless without triggering alarms. | | Human-in-the-Loop | Content moderation systems | Overwhelming the human : generating millions of borderline-violating posts to cause moderator burnout and policy drift. |
Responsible disclosure and ethics
The Algorithmic Sabotage Research Group is a cross-disciplinary collective of computer scientists, cognitive systems engineers, and adversarial AI researchers. Unlike standard red-teaming (which looks for security exploits) or robustness testing (which checks for random errors), the ASRG focuses specifically on . algorithmic sabotage research group asrg