Model Auditing

Model auditing is the independent examination of an AI system to check how it behaves, what risks it carries, and whether it meets stated requirements. Audits can cover things like bias, accuracy, safety testing, data handling, and documentation, and they are often performed by third parties or internal review teams separate from the people who built the model. For compliance and risk leaders, auditing matters because regulators and frameworks such as the EU AI Act increasingly expect organizations to show evidence of this kind of review, not just claim their systems are safe and fair.

1 item

ResearchGlobal2026-11-03

AI-to-AI Weight Editing Breaks the 'Tested Model' Governance Assumption

MIT CSAIL researchers have demonstrated that GPT-6 Astra can directly rewrite the internal parameters of other AI models, boosting benchmark performance without conventional retraining. The finding undermines the core governance assumption that a model's behavior is fixed and verifiable after pre-deployment review. Organizations relying on one-time conformity assessments or static model audits face a structural gap in their AI assurance programs.

GPT-6 Astra Qwen3.5-0.8B OpenAI MIT CSAIL model integrity weight manipulation AI safety model auditing