## Thomson Reuters builds its own legal AI instead of renting from OpenAI
Thomson Reuters builds its own legal AI instead of renting from OpenAI
Thomson Reuters launched "Thomson," its first in-house language model for legal work — built on Alibaba's open Qwen rather than fine-tuning a frontier model from OpenAI or Anthropic, a bet on owning AI instead of renting it.
The company says it spent about $40 million over two years on the project (the final training run alone came to roughly $450,000). Thomson starts from Alibaba's Qwen — most recently Qwen3.5-397B — which the company first retrained, with Imperial College, for safety, ethics and political neutrality in an intermediate version it calls "Snowdon." From there it pre-trained on its own content across Westlaw, Practical Law, Checkpoint and Reuters, post-trained with hundreds of domain experts, and ran agentic reinforcement learning inside its own tool environments. Critically, less than 10 percent of that content has been used so far, so the ceiling is untapped.
The benchmark picture is more honest than the company's "among the world's best" framing. On Stanford LegalBench, Thomson (0.823) trails Gemini 3.1 Pro and GPT-5.5; it sits just behind Opus 4.8 on the Harvey Legal Agent Benchmark but leads on instruction-following and the PrBench Legal set. On web access alone it scores 0.53 on the firm's factual-accuracy deep-research test versus GPT-5.4's 0.65 — and it only edges past GPT-5.4 (0.83 to 0.82) when it can reach Thomson Reuters content. In other words, proprietary data access does almost as much work as the specialized training, and the numbers haven't been independently validated yet. The comparison is also somewhat lopsided: Thomson runs with test-time scaling while GPT-5.5 was tested without a reasoning mode.
The strategy, as CTO Joel Hron frames it, is "renting a house versus buying a house": third-party fine-tuning tends to degrade general capability, locks a buyer into a provider's roadmap and inference pricing, and hands every bit of captured expertise back to that provider. With an in-house model, "you are building equity in something that you own for the long-term," Hron says, and that compounds as each expert review during a product update becomes training data. Thomson launches in the Tabular Analysis feature of CoCounsel Legal, with a smaller open-weight version coming to Hugging Face under a non-commercial license. Taken honestly, this is a proof case that open-source now trails the frontier labs by only months, and that a focused $40 million is enough for a genuinely competitive specialized model — for companies that own proprietary data, domain experts and a way to measure output objectively.
Fake Codex installer tricks Mac users into pasting malware
Attackers are weaponizing the AI-coding-tool rush — Cato Networks' threat-research team detailed a campaign built around a fake OpenAI Codex installer that lures Mac users into pasting malware into Terminal.
The lure is a sponsored Google result for queries like "codex macos download" that appears above OpenAI's own listing. Clicking it lands on a Google Sites page that copies the Codex download portal; the page itself carries no malicious code, with attacker content loading inside an iframe routed through a Google proxy, so operators can swap payloads without touching the visible page. The victim walks through opening Terminal and pasting a command that starts with a plausible install string, then decodes a URL and pipes a remotely fetched script into the shell. Cato ties the delivery framework to Atomic macOS Stealer (AMOS), the commodity infostealer, with matching loader structure, telemetry calls and staging paths.
What makes this notable is how defensively clever the delivery is: the obvious page path returns a harmless product page while the real ClickFix lure sits behind a near-identical route, non-macOS visitors get benign content, and the loader strips the download's quarantine metadata so macOS doesn't show its usual warning before running it. As Cato notes, "no single stage reliably exposes the attack" — detection depends on correlating search placement, embedded content, Terminal execution and outbound traffic. It's also the latest in a pattern of malware dressed up as AI developer tooling this year, and a reminder that the fastest-adopting corner of the AI buildout is exactly where the social engineering sits.
Do you think owning a model — data and all — is the real moat in enterprise AI, or does renting the frontier still win? Tell us in the comments. Sources: The Decoder · SiliconANGLE · Thomson Reuters · SiliconANGLE — fake Codex installer
