KPMG and Hitachi Announce Anthropic Partnerships in Quick Succession — the Enterprise AI Adoption Wave Accelerates
On May 19, KPMG, a global professional services firm, and Japan's Hitachi, Ltd. both announced strategic partnerships with Anthropic in quick succession.
KPMG is a major firm delivering audit, tax, legal, and advisory services across 138 countries and territories. Under the new partnership, KPMG will integrate Claude into its workflow platform, Digital Gateway, starting with new tools for tax and legal clients. More than 276,000 employees worldwide will gain access to Claude. Anthropic has also named KPMG a priority partner in the private equity space, and the two companies will co-develop Claude-powered products for PE portfolio companies.
On the same day, Hitachi, Ltd. announced a strategic partnership with Anthropic as well. The company plans to deploy advanced AI such as Claude across all business processes of its roughly 290,000-person group, and to extend it to its social infrastructure solution, HMAX by Hitachi. Anthropic had also announced a collaboration with NEC in April, steadily building out partnerships with major Japanese IT companies.
KPMG Global Chair and CEO Bill Thomas emphasized that the firm will "prioritize responsible AI and scale with a focus on security, trust, and governance." Large-scale enterprise AI deployments unfolding simultaneously across the West and Japan are a symbolic moment that signals the maturing of the enterprise AI market.
Cloudflare Discovers Exploit Chains That Prior Models Missed, Using Mythos
Cloudflare reported that it tested Anthropic's security-specialized model, Mythos Preview, across more than 50 of its own code repositories and discovered exploit chains — chained attack paths — that previous frontier models had missed.
The evaluation was part of a Cloudflare security validation project called Project Glasswing, designed to gauge Mythos's capabilities on real production codebases. The fact that Cloudflare — one of the world's largest infrastructure companies — validated the model on its own code makes this a significant piece of evidence for Mythos's practical effectiveness.
Unlike previously reported vulnerability discoveries in financial infrastructure, these results demonstrate performance in the web infrastructure domain. The effectiveness of specialized AI models in security is now being confirmed across multiple industries.
ByteDance Releases "Lance" — a 3B-Parameter Model Unifying Image and Video Understanding, Generation, and Editing
ByteDance has open-sourced Lance, a model that handles image understanding, image generation, image editing, video generation, and video understanding within a single framework — with just 3B parameters.
Most notable is its low training cost: the model was reportedly trained from scratch using 128 A100 GPUs. It delivers strong performance for the 3B class across image generation, editing, and video generation benchmarks.
While "understanding" and "generation" have typically been handled by separate models, Lance takes a unified approach — a sign that the trend toward lighter, more integrated multimodal AI is reaching a practical level.
CHI-Bench: Healthcare AI Agents Are Still Far from Production-Ready
CHI-Bench, published on Hugging Face Daily Papers, is the first full-scale benchmark to test how much of healthcare's end-to-end workflows AI agents can automate.
It builds high-fidelity simulators on MCP servers for three healthcare domains: prior authorization, utilization review, and care management. Each trial is a long-horizon task spanning 60–80 steps and 4–6 clinical process stages, involving more than 200 MCP tools and 1,279 documents of operational manuals.
Of the 30 frontier agents tested, the top performer was Anthropic's Claude Code (Opus 4.6) with 28% pass@1, followed by OpenAI's Codex (GPT-5.5) at 21%. No agent could reliably clear 20% or more when running the same cases three times.
The results underscore that despite improving agent performance, healthcare work — with its complex policies and audit-trail requirements — is not yet at a stage where it can be delegated without human review.
HRM-Text 1B: A Reasoning-Focused Model Trained for About $1,000 Outperforms Larger Models
HRM-Text 1B, released by Sapient Intelligence, packs just 1B parameters yet scored 56.2 on the MATH mathematical reasoning benchmark (beating Llama3.2 3B's 48.0) and 82.2 on DROP.
Its training setup is remarkably efficient: 16 GPUs, 1.9 days, 40B tokens of training data, and a budget of roughly $1,000. The model reportedly trains on 100–900× less data than models with comparable performance.
True to its name — Hierarchical Reasoning Model — it is especially strong on reasoning tasks, though it still lags behind larger models on general knowledge (MMLU). As an approach to maximizing a specific capability with limited resources, it is drawing significant attention in the local AI community.
UK Survey: Fear About AI's Future Outweighs Hope
According to a survey by King's College London, one in five UK citizens believes AI will cause civil unrest. On the future of AI and work, more people report feeling fear than hope.
While AI's technical progress has been remarkable, public acceptance has not necessarily kept pace. Anxiety about the impact on jobs remains deep-rooted, confirming once again the gap between the tech industry's enthusiasm and broader social concern.
---