join.dev
AI news · 10 stories
Epoch AI demonstrates Firefox patch-gap exploits with GLM 5.3 · Google reports offensive AI-agent frameworks in real intrusions · CrowdStrike details Agent Graph for tracing AI endpoint activity
Epoch AI’s October 29 report says GLM 5.3 built file-reading exploits for six Firefox releases using public but unshipped fixes. Mozilla patched an additional vulnerability used in the attacks; the study highlights risks in the interval between publishing fixes and delivering updates.
Read at Epoch AI
Google Threat Intelligence Group and Mandiant reported on October 30 that attackers are abusing frameworks including ARTEX, Hexstrike and CAIRN across multiple stages of intrusions. Their observations show agent automation moving beyond reconnaissance into operations that give defenders less time to respond.
Read at Google Cloud
CrowdStrike’s October 30 explanation of Agent Graph links prompts, tools and models to endpoint processes, files and detections in Falcon Guardian. The connected view helps security teams trace what supported agents actually did and investigate the actions behind alerts.
Read at CrowdStrike
IBM released the backward-compatible Granite Switch 4.1.1 patch on October 29, adding speech recognition across ten languages with punctuation and an updated decision adapter. The base model stays unchanged, illustrating how modular adapters can deliver improvements without rebuilding the entire model.
Read at Hugging Face
Hugging Face and research partners announced Project ProtoStar on October 29 to make publicly funded datasets, models and code easier to reuse in AI research. Its first open challenge, hosted by Carnegie Mellon University, starts November 2 and focuses on urban climate prediction.
Read at Hugging Face
AnySpeechRecognizer and AnyDecisionModel were announced on October 29 with Swift APIs for speech recognition and structured decisions using a choice of models. Both support downloading open models for on-device use, expanding local AI options for Apple-platform developers.
Read at Hugging Face
Microsoft’s October 30 research post describes an automated platform that builds synthetic organizations, work histories and queries to evaluate Copilot without customer data. Industry-specific environments help test realistic workflows and reveal issues that generic benchmarks can miss.
Read at Microsoft Research
Wabi announced $55 million in funding led by IVP at a $1.25 billion valuation on October 29. The financing supports an assistant combining personalized mini-apps and shared spaces for everyday planning, extending consumer AI beyond individual chat sessions.
Read at PR Newswire
Epoch AI’s October 29 analysis estimates that the computing capacity of AI chips sold each quarter has grown about 2.5 times annually since 2024. The estimate tracks chip sales rather than operational data centers, providing a measure of infrastructure expansion with deployment delays still to account for.
Read at Epoch AI
Epoch AI’s October 30 analysis finds that 17% of September biology preprints acknowledged LLM use, approaching the 19% citing specialized AI tools. These voluntary disclosures indicate changing research practices, but do not measure all actual AI use or establish research quality.
Read at Epoch AI