iFeeling Daily
Daily curated AI insights you can't miss.
Scaling Linux Kernel Vulnerability Research to 948 LPE Exploits
One research prototype scanned 7% of the Linux kernel and built 948 local privilege escalation exploits from 1,724 validated vulnerabilities.
Model Integration Contracts: Catching Regressions Before They Reach Main
Transformers builds small fixture models and Model Integration Contracts to catch regressions like the UMT5 SDPA bug before they reach main.
Pando: OpenAI’s Git Backend for Agent-Scale Source Control
OpenAI built Pando, a new Git backend that scales concurrent pushes and fetches for agent-driven development, cutting median push latency by nearly 5x.
LLMJacking: How Attackers Monetize Compromised Cloud LLM Access
Attackers hijack enterprise cloud LLM access via stolen credentials and resell it through underground LLM Mixers, racking up millions in fraudulent charges.
Detecting Frontier Model Misalignment: OpenAI’s Multilayer Monitoring System
OpenAI describes its multilayer misalignment monitoring system, which would have flagged the Hugging Face incident over 30 hours before the actual breach.
How Google Ships Agent Skills Like Products: Versioning, Evals, and AI Review
Google Agent Skills ships skills like products: versioned, validated, evaluated, and human-approved before merge, with AI review agents and CI gates.
Ai2 Finds LLM Novelty Judges Are Unstable on AI-Generated Research Ideas
Small prompt changes swing LLM novelty-judge accuracy by over 50 points on Ai2's new benchmark. Simple baselines beat purpose-built evaluators.
Securing OpenAI’s Frontier Model Training Environments
OpenAI cut off live internet access for frontier training after containment incidents. It built an offline world and red-teamed with the models themselves.
Scaling OpenAI’s Habitat Storage Platform on Azure Cosmos DB
OpenAI's Habitat storage platform now serves over 1 billion users, using multi-tenant Azure Cosmos DB containers and joint engineering with Azure.