Claude adds ready-made tools for legal work
Lawyers keep the final say on source-linked results. Plus, Google puts cyber defenders first in its next model rollout.
By Saeed Ezzati8 min read
The audio edition
Listen to this newsletter
0:004:39
Read transcript
Anthropic is turning legal work in Claude into reusable, inspectable workflows. Its new gallery offers ready-made tools for tasks like checking citations, comparing rules across jurisdictions, tracking matters and monitoring upcoming regulations. Each opens with an example; a team can then add its own documents or connect research sources, and adapt or share the result. Anthropic calls these interactive tools artifacts. The point is not to hand legal judgment to a model. Claude makes a first pass, with outputs linked to their sources, and a lawyer reviews the work. Anthropic says the citation checker flags quotations that do not match, incorrect page numbers and sources that do not support a claim. Contract-review tables extract requested terms and link back to the exact language. Another tool compares contracts with a company’s standard terms, flags exceptions and routes them to the person responsible for approval. Other examples turn jurisdiction research into a sourced comparison table, organize open matters by owner and deadline, or place relevant regulatory changes on a timeline. Teams can ask Claude to tailor an artifact and connect outside systems rather than manually uploading everything. Anthropic says it has added more than 60 legal connectors since the product launched in May—not all with this gallery. Sharing settings are intended to support privacy and ethical walls between teams. The gallery also includes partner-built tools, including Harvey Intake for routing legal requests, and nine tools aimed at access to justice. Anthropic says five were built by nonprofit and public-interest groups it supported with Claude credits. The broader shift is from asking a model for an answer to building a repeatable process around its first pass—and making the review handoff visible. That same emphasis on a defined handoff is shaping Google’s next model launch. Google says Gemini 4 Carbon is built for complex engineering tasks that can run for days, and selected cyber defenders are due to receive access first this week. Trusted partners will follow in stages; Google has set no firm date for developer, business or consumer access while it gathers feedback and refines safeguards. Google reports scores of 88.2% for finding vulnerabilities, 54.8% for validating whether they can be exploited, and 71% for patching them. The cybersecurity evaluations ran without guardrails, an important caveat. The company also says its internal teams have seen Carbon work through extended coding tasks, and reports improvements in security and resistance to prompt injection. Those performance and safety claims come from Google’s announcement, not independent testing. The practical question is whether the staged rollout can produce useful defender feedback without treating benchmark results as a guarantee of safe behavior. And in a very different kind of review workflow, Italian police used visual-matching AI to help recover a stolen painting. New BBC reporting says officers photographed works at an antiques fair in Parma in October 2024. The system flagged a portrait that appeared to show a woman holding a scroll; investigators identified it as Saint Agnese, stolen in 2004, whose original lamb, halo and palm branch had been overpainted or removed. Police seized it the next day for examination. The system compares visual features such as color, outlines and composition, rather than relying only on descriptions. But it generates leads, not authentication: incomplete records, poor images and copies can complicate a match, and investigators still have to establish what they have found. The case shows the value of visual search when a changed work re-enters a market—not that AI can settle provenance on its own. That distinction between a useful tool and an accountable decision is also at issue near xAI’s Mississippi power plant. The Guardian reports that an xAI subsidiary is offering to buy nearby homes, with a condition barring sellers from joining a residents’ noise lawsuit. County records show five purchases since October 9. A realtor reportedly described a larger target, but xAI has not confirmed those figures or the scale of its plans. Residents allege persistent noise and sleep disruption; xAI denies the lawsuit’s claims. The company has changed power output and built sound-dampening walls, while the city says its monitoring has not found a breach of the local noise limit. A state agreement calls for replacing temporary turbines with a permanent gas-fired plant by summer 2027, which the Guardian reports should reduce noise. Across these stories, the useful thing to watch is where the first pass ends and responsibility begins: a lawyer’s review, a defender’s testing, an investigator’s confirmation, or a company’s concrete remedy for affected neighbors.
