Humanbound ships v2.13.0 of its open-source adversarial testing engine, SDK and CLI for AI agents, licensed Apache-2.0. The company says hb guardrails turns failed test findings into deployable firewall rules, and hb posture scores an agent's security from 0 to 100. Runs locally or against the Humanbound Platform.
Anthropic launches OSS Scanner, an opt-in service that sends model-generated vulnerability reports, with no human triage, to eligible open-source projects at no cost. Anthropic says six months of scanning produced over 29,000 candidate vulnerabilities, of which only about 6,000 were reviewed. In testing, 85 of 97 critical findings across 48 projects met its disclosure bar.
Goodfire releases probe-based cyber monitors for Kimi K3 and GLM 5.3 that, the company says, match an LLM judge's 93% recall at 5.5% false positives for about $65 to $185 per million turns, roughly 50x cheaper. Goodfire reports no throughput loss. FAR.AI's red-teaming found 0 universal jailbreaks across 140 strategies.
Hugging Face's HuggingFaceBio team releases Carbon-A, a 1.2-billion-parameter model that predicts protein-coding genes from DNA, and the Carbon Annotation Database of 566 million gene candidates across 22,617 species. The team reports a 0.944 nucleotide F1 across 42 benchmark genomes. Iso-Seq tests in cat, hamster, chicken and Arabidopsis support some predictions missing from RefSeq.
Liquid AI releases two open-weight decision models on Hugging Face: d1-3B and d1-omni-600M. The company says d1-3B scores 48.57 on its Decision Index v0.2.1, ahead of every model under 10B. It reports 8 ms answers on an RTX 4090 and 50 ms on a Jetson Orin Nano. d1-omni-600M handles text, image and audio.
Ai2's research on Bolmo, a method that converts subword language models to byte-level models, appears in Nature. The institute releases Bwen 8B and Blama 8B, built from Qwen 3 8B and Llama 3 8B, with Bwen 8B beating Bolmo 7B across its evaluation suite. Stage 1 checkpoints keep the original model frozen to speed experimentation.
Perplexity has released pplx-embed-v2-late, two late-interaction embedding models of 0.6B and 9B parameters that handle text-to-text and text-to-image retrieval in one shared space. Both checkpoints are on Hugging Face and work with sentence-transformers and transformers. The company says the smaller model matches systems with five times its active parameters on the ViDoRe (V3) benchmark.
NVIDIA ships version 1.0 of its AI Cluster Runtime, an open-source tool that pins compatible component versions for GPU-accelerated Kubernetes clusters. The release locks the CLI, REST API, Go SDK, bundle layout, and artifact schemas under a compatibility contract with merge-blocking checks. More than 100 people contribute, and integrations exist with Pulumi Labs and Mirantis's k0rdent.
Google releases EmbeddingGemma 2, an open-source 740-million-parameter on-device model that handles text, code, images, video and audio, and which Google claims is best-in-class for embedded models under 1 billion parameters. New York startup Reflection launches Beam, a 500-billion-parameter open-source model for coding, reasoning and agentic work. Reflection says Beam runs frontier reasoning up to four times cheaper than GLM-5.2.
Mistral launched a public preview of Mistral Large 4, called le Chonk, a 1-trillion-parameter open-weight model for agentic tasks. The company claims it matches the best closed models and trained it on 4,000 NVIDIA Grace Blackwell GPUs. The weights arrive October 27, and co-founder Guillaume Lample says open weights protect cybersecurity tools from sudden deprecation.