Here's your daily roundup of the most relevant AI and ML news for September 21, 2026. Today's digest includes 1 security-focused story. We're also covering 5 research developments. Click through to read the full articles from our curated sources.
Security & Safety
1. Claude Opus 5 Helped Researchers Take Over OpenAI Staff Accounts via Chained Flaws
Three researchers at the security firm Hacktron used Anthropic's Claude Opus 5 to chain two flaws and take over the ChatGPT and Codex accounts of several OpenAI employees, then reach an internal OpenAI code repository.
The chain began with a bug in the software that runs OpenAI's publi...
Source: The Hacker News (Security) | 1 day ago
Research & Papers
2. HE-Guardrail: A Homomorphic Guardrail Against Jailbreak Attacks for Encrypted Large Language Model Inference
arXiv:2609.21484v1 Announce Type: cross Abstract: Homomorphic encryption (HE) has emerged as a promising approach to privacy-preserving machine learning (PPML), enabling computation directly over encrypted data. In HE-based PPML, a client submits an encrypted input to the server, which evaluates...
Source: arXiv - AI | 10 hours ago
3. Taming the Adversary: A Cost-to-Disturbance Ratio Approach to Adversarial Reinforcement Learning
arXiv:2603.12110v2 Announce Type: replace Abstract: Reinforcement learning (RL) policies trained in simulation often degrade once deployed on real systems, where the controller must reject external disturbances that were never encountered in simulation. Robust RL addresses this by exposing the c...
Source: arXiv - Machine Learning | 10 hours ago
4. Near-Optimal Primal-Dual Algorithm for Learning Linear Mixture CMDPs with Adversarial Rewards
arXiv:2603.27884v2 Announce Type: replace Abstract: We study safe reinforcement learning in finite-horizon linear mixture constrained Markov decision processes (CMDPs) with adversarial rewards under full-information feedback and an unknown transition kernel. We propose a primal-dual policy optim...
Source: arXiv - Machine Learning | 10 hours ago
5. How Much of a Real Workload Can LLM-Generated GPU Kernels Actually Reach?
arXiv:2609.21058v1 Announce Type: cross Abstract: Language models can now write GPU kernels that outperform PyTorch. We evaluate five model configurations on KernelBench level 1 and find that a frontier model produces correct kernels for 91.1% of problems and independently verified speedups on 2...
Source: arXiv - AI | 10 hours ago
6. An Introduction to Compression-Based Machine Learning
arXiv:2609.21309v1 Announce Type: new Abstract: Any lossless compression algorithm (like gzip) may be converted into a machine learning method, via either Normalized Compression Distance or the Minimum Description Length principle. Any auto-regressive model may be converted into a lossless compr...
Source: arXiv - Machine Learning | 10 hours ago
Industry News
7. AI safety conversations have gotten unbelievable
This week two conversations about AI safety went viral that demonstrate just how hard it is to discern AI fact from fiction.
Source: TechCrunch - AI | 1 day ago
Tech & Development
8. The Roadmap to Mastering LLM Inference Optimization
Article URL: https://machinelearningmastery.com/the-roadmap-to-mastering-llm-inference-optimization/ Comments URL: https://news.ycombinator.com/item?id=49786379 Points: 1
Comments: 0
Source: Hacker News - AI | 1 hours ago
About This Digest
This digest is automatically curated from leading AI and tech news sources, filtered for relevance to AI security and the ML ecosystem. Stories are scored and ranked based on their relevance to model security, supply chain safety, and the broader AI landscape.
Want to see how your favorite models score on security? Check our model dashboard for trust scores on the top 500 HuggingFace models.