AI Daily — June 3, 2026
2026-06-03
anthropic.com
What we learned mapping a year's worth of AI-enabled cyber threats
Anthropic published findings from a year-long analysis of AI-enabled cyber threats mapped to the MITRE ATT&CK framework. The report characterizes how threat actors are incorporating LLMs into attack workflows across reconnaissance, phishing, code generation, and vulnerability exploitation. This is a concrete empirical dataset on real-world offensive AI use, relevant to defenders building detection and response strategies.
anthropic.com
Expanding Project Glasswing
Anthropic announced an expansion of Project Glasswing, its initiative focused on AI safety and interpretability research partnerships. The expansion broadens the program's scope and external collaborations, signaling continued investment in third-party safety research alongside internal efforts. Details on new partners and specific research directions are outlined in the announcement.
openai.com
Codex for every role, tool, and workflow
OpenAI announced new Codex integrations including plugins, site connectors, and annotation tools targeting non-engineering roles such as analysts, marketers, designers, and investors. The rollout extends Codex beyond software development into broader enterprise workflows, positioning it as a general-purpose agentic coding layer. This represents a deliberate effort to widen Codex's addressable market beyond developers.
openai.com
Travelers deploys AI-powered claims countrywide with OpenAI
Travelers Insurance has deployed a nationwide AI-powered Claim Assistant built on OpenAI's models to guide customers through the claims filing process and provide 24/7 support. The system is designed to scale operations during peak demand periods such as after natural disasters. This is a notable production deployment of LLM-based agents in a regulated, high-stakes insurance context.
huggingface.co
Holo3.1: Fast & Local Computer Use Agents
H Company released Holo3.1, a model designed for computer use agents that runs locally with a focus on speed and on-device execution. The release targets use cases where cloud-dependent computer control agents are impractical due to latency, privacy, or cost constraints. Local computer use agents remain a technically challenging frontier, making this a notable open release in the space.
arxiv.org
Do Value Vectors in Deep Layers Need Context from the Residual Stream?
Researchers find that in deep transformer layers, model performance improves when value vectors are computed context-free—without drawing on the residual stream—suggesting these layers primarily function to preserve original token information rather than integrate contextual signals. The context-dependent component adds little incremental benchmark performance when the context-free vector is available. This has architectural implications for transformer design and efficiency, as context-free value vectors can be stored as sparse parameters.