Anthropic launches Cyber Mission: free scans for open-source projects and frontier Claude for critical infrastructure
On 2026-10-08 Anthropic announced the Anthropic Cyber Mission, a long-term program to give defenders tools and research. It has two starting points: a Critical Infrastructure Defense Program that pairs frontier Claude models, on-site engineers and threat research with providers protecting operational technology, and OSS Scanner, which gives enrolled open-source projects free periodic scans from Anthropic's strongest models (Anthropic expects a true-positive rate above 90% for its reports, a vendor expectation not independently measured). Eleven founding partners include CrowdStrike, Palo Alto Networks, Dragos, Accenture, Deloitte and Rockwell Automation; Anthropic has also funded open-source foundations including the Python and Apache Software Foundations, and says a government cyber-defense program is on offer to more than half of US states.
Sources & depth
TRL 1.15 makes a fused LM head the default, allowing much longer training sequences on the same GPU
Hugging Face's TRL v1.15.0 turns on a fused LM head by default for SFT, DPO, KTO, GRPO, RLOO and distillation trainers. A Triton kernel computes per-token log-probs and entropy without ever building the full batch-by-sequence-by-vocabulary logits tensor, which the author reports allowed KTO sequences almost seven times longer on the same GPU in a gemma-3-1b test (a single author benchmark). The release also adds selective activation checkpointing for SFT and assistant-only loss on vision datasets, but drops Python 3.10, removes the experimental MiniLLM trainer and deprecates use_liger_kernel in four trainers ahead of v2.0.0.
Sources & depth
- github.comv1.15.02026-10-09 00:56 IST
Ollama v0.40.2 upgrades previously downloaded models in the background
Ollama v0.40.2 upgrades models downloaded with earlier versions the first time they are run, for better performance and llama.cpp compatibility. The originals are kept as backups and take disk space until a future release removes them, and going back to a version older than 0.40 means re-pulling those models. The release also stops ollama list from showing duplicate entries and makes ollama launch claude use the model's full context length.
Sources & depth
OpenAI says it disrupted two AI-enabled influence operations built on false-front journalists and a think tank
OpenAI reports it disrupted two influence operations that used AI to run fake journalists and a think tank to spread geopolitical messaging. The page was unfetchable, so scale, origin and methods are link only.
Sources & depth