Red Hat benchmarks show small classifiers rival large LLMs for prompt injection
A 200M-parameter model matched a 35B LLM judge on prompt injection accuracy while running significantly faster, according to new Red Hat benchmarks.
RAG, evaluations, inference, fine-tuning and shipping LLM features to production.
A 200M-parameter model matched a 35B LLM judge on prompt injection accuracy while running significantly faster, according to new Red Hat benchmarks.
OpenAI introduces textGrain, an opt-in statistical watermarking system for API text. Detection rates vary by content type and editing, with code posing significant challenges.
Santhosh Sundar outlines how to build compliance systems that enforce rules at the point of decision using deterministic code for hard limits and agents for interpretation.
A new open-source tool uses Ollama and Nimble to detect prompt injections with structured data instead of generative text, keeping security checks local.
Speed of light modeling shows that FSDP creates a communication bottleneck on InfiniBand for DeepSeek-V3, making pipeline parallelism the necessary choice for efficient training.
Amazon demonstrates how multi-turn RL fine-tunes a Qwen3.6-27B search agent, cutting failure rates from 22% to under 1% while improving retrieval quality.
A technical workflow using NVIDIA NeMo reduces word error rates for Najdi and Hijazi speech by combining weighted replay mixing and partial encoder unfreezing.
Graph RAG adds explicit relationship evidence to vector search, helping agents answer multi-hop questions about dependencies and policies without hallucinating connections.
A technical guide details how to configure cross-account SigV4, workspace-scoped API keys, and OIDC federation for secure LLM inference across diverse environments.
Cloudflare’s AI Search is now generally available, adding native image embeddings, OCR for PDFs, and support for larger files. Billing starts November 1, 2026.
Cloudflare updated its Containers infrastructure to support runtime image selection and filesystem snapshots, reducing startup times by over six times for AI agent workloads.
OpenAI introduced a new Decisions API at Dev Day, offering fast, low-cost classification similar to TypeSafe AI's Jev model for securing AI agents.
Cohere released Embed 5, allowing developers to index with a high-quality Pro model and query with a faster Fast model in the same vector space.
NVIDIA demonstrates an end-to-end workflow for serving Hierarchical Sequential Transduction Unit models using PyTorch AOTI, FlexKV caching, and Dynamo-Triton to reduce inference latency.
NVIDIA releases VSS Blueprint 3.3, introducing Adaptive Efficient Video Sampling to reduce token usage and a Build Vision Agent skill that composes multi-workflow deployments from a single prompt.