AI Daily — July 16, 2026
2026-07-16
openai.com
GPT-Red: Unlocking Self-Improvement for Robustness
OpenAI has released details on GPT-Red, an automated red-teaming system that uses self-play to generate adversarial prompts and improve model robustness against jailbreaks, prompt injection, and alignment failures. The system was used during training of GPT-5.6, with OpenAI claiming it made that release significantly more resistant to cyberattack-style exploits. The self-play loop — where GPT-Red iteratively finds weaknesses that are then patched into successor models — represents a scalable alternative to manual red-teaming.
openai.com
The US is advancing AI safety through state and federal action
OpenAI published a policy position advocating a 'reverse federalism' model for AI governance, where state-level AI safety legislation serves as a testbed and foundation for a eventual national framework rather than being preempted by federal action. The post argues that democratic, bottom-up regulatory development is preferable to top-down federal preemption. This is notable given ongoing Congressional debates about whether federal AI law should override state bills.
rss.arxiv.org
OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets
OriginBlame (ob) is a data provenance system that tracks author identity at the record and token level through entire data processing pipelines, enabling precise forget-set generation for machine unlearning when contributors request removal. Tested on 219,555 Wikipedia pages, it reduces dataset-level over-deletion from 101x to 1.3x compared to coarser file-level systems. Integration overhead is modest — 1.3–4.0% throughput penalty on HuggingFace pipelines and 2.1–19.0% on Datatrove — making it practically deployable for compliance with right-to-erasure requests.
rss.arxiv.org
Targeted Recovery of Weight-Space Mechanisms From Neural Networks
Targeted Parameter Decomposition (tPD) extends mechanistic interpretability by decomposing only the network components that process specific inputs of interest, using a high-rank catch-all component to absorb all non-target computation. This cuts the FLOP cost of circuit extraction to ~7% of full decomposition while still recovering mechanistically faithful, reproducible circuits on transformer language models trained on The Pile. The approach makes weight-space circuit analysis tractable for larger models where full parameter decomposition was previously cost-prohibitive.