Linklaters debuts applied AI practice for bespoke legal tools

TLDR: A Magic Circle law firm is productizing in-house agents, Mozilla is quietly shipping AI-found Firefox bugs, and AI is creeping into hurricane forecasts.

Linklaters creates Applied Intelligence practice for case-specific AI

Global law firm Linklaters has launched an Applied Intelligence practice that builds matter specific AI tools using large language models from Google, Anthropic, and OpenAI, plus Databricks and APIs across its data stack. The team is already using this to triage large volumes of class action claims and similar workflows, according to a report in Law.com as of 2026-05-10.

For anyone building production agents in regulated domains, this is a clear signal that top tier firms are moving from “AI pilot” to “internal product team.” Linklaters is effectively standing up an AI platform group that can assemble bespoke workflows per matter using retrieval augmented generation (RAG), enrichment, and internal data sources. That means tighter integration with document management, billing, and risk systems, not just generic chatbots.

The interesting part to watch is governance: how they handle provenance, review, and client consent when an agent touches case strategy or discovery at scale. If this model works, you should expect every large firm or enterprise to copy the “applied intelligence” template within a year.

Read more →

---

Mozilla uses Anthropic Mythos to fix 271 Firefox vulnerabilities

Mozilla reports that Anthropic’s Claude Mythos Preview helped identify and fix 271 vulnerabilities in Firefox, all landing in Firefox 150, with “almost no false positives” as of 2026-05-10. Earlier work using Anthropic Claude Opus 4.6 reportedly led to fixes for 22 security sensitive bugs, according to Mozilla’s own blog and follow up coverage.

This is notable because it is not a toy eval. It is a shipping browser with a long legacy codebase, and Mythos appears to have produced actionable reports at scale that engineers trusted. For security teams and infra engineers, this suggests a concrete pattern: pair specialized LLM variants with structured review workflows instead of generic “copilot for security” pitches.

The catch: we do not yet have granular precision and recall metrics or reproducible benchmarks, only Mozilla’s internal assessment and anecdotes from engineers. Still, if a major OSS project keeps this in the toolchain, expect similar “LLM as static analyzer” patterns to spread into your CI within the year.

Read more →

---

AI models join National Hurricane Center forecasting for Texas

The Houston Chronicle reports that artificial intelligence based hurricane forecasting is being integrated into National Hurricane Center workflows, promising earlier and more precise alerts for Texas Gulf Coast residents as of 2026-05-10. These AI models analyze historical storm tracks and real time data to refine cone forecasts and intensity predictions.

For anyone building decision support agents around weather, insurance, logistics, or utilities, this is a sign that AI outputs will become first class signals in emergency operations. You can design agents that reason over both traditional numerical weather prediction models and AI ensembles, then expose scenario planning to operators or citizens.

Worth noting: forecasters still treat AI as an augmentation layer, not a replacement, and there are no long term skill scores published yet. If you are wiring this data into products, keep humans firmly in the loop and make uncertainty legible to end users.

Read more →

---

Quick Hits

• OncoAgent: A Dual-Tier Multi-Agent Framework for Privacy-Preserving Oncology Clinical Decision Support This Hugging Face blog details OncoAgent, an open source LangGraph based multi agent system that uses retrieval augmented generation (RAG) and QLoRA tuned models on AMD hardware to assist oncology decisions while keeping patient data local as of 2026-05-10.
• Show HN: GETadb.com – every GET request creates a DB GETadb gives each GET request a fresh backend with database, sync engine, auth, presence, and streams so browser or agent clients can spin up full stack apps without credentials; it even serves different content to curl for agent friendliness.
• Using Claude Code: The Unreasonable Effectiveness of HTML Thariq Shihipar and Simon Willison argue that you should ask Claude Code for HTML output instead of Markdown, since HTML lets agents express layout, semantics, and interactions more precisely, with concrete prompt examples.
• Teaching Claude Why Anthropic’s latest research note explores training methods that push Claude to explain its reasoning and internalize causal structure, which has implications for debugging and tool using agents that must justify actions.
• Chat SDK adds web, Messenger adapters and conversation history Vercel’s AI Chat SDK now supports a browser web adapter, a Messenger adapter, and cross platform conversation history primitives so your bot can keep one transcript as users switch channels. Also covered by: Messenger adapter, conversation history.