News You Can Use

Spectre is an autonomous company agent that operates without a human prompt. It monitors the business and acts on incidents, bug reports, customer feedback, and internal signals. It sits on what Harvey calls a "law firm world model" - a live picture of what is happening inside an organisation and what needs to happen next. The same week, Harvey published research on "harness engineering" - agents teaching themselves to do legal work through iterative self-improvement. On internal benchmarks, average scores jumped from 40.8% to 87.7% with no model update and no manual prompt engineering. Seven of twelve tasks exceeded 90%. The agents generated their own legal-specific tools along the way: cross-document playbooks, validation checkpoints, structured fact sheets. The headline is self-improvement. Given good examples and a rubric, the agents got meaningfully better at legal work by themselves.

Capacity is no longer the constraint. The technology can do the work. The question for innovation leaders is whether the firm has the people with the judgement to decide what needs to be done and what good looks like. Given examples and a rubric, agents now auto-generate their own toolkits and improve without a model update. The compounding advantage lives with the firm that owns the examples and the rubric. Static prompt libraries are already behind. And with Spectre removing the human trigger, the bottleneck shifts to review, prioritisation, and coordination, all of which is judgement work. Partners have a different job. Juniors have a different apprenticeship. The question worth answering is how to rebuild supervision, judgement, and pricing around a workforce that can produce more than the firm can currently absorb.

Microsoft 365 Copilot's Researcher agent can now use OpenAI's GPT to draft a response, then have Anthropic's Claude check it for accuracy and citations before finalising. Two features make this work: Critique (one model checks another's output) and Council (multiple models collaborate). It ships as part of Microsoft 365 E7 "Frontier Suite". Three days later, Anthropic launched Claude for Word in beta with legal contract review as the flagship use case. It ships through Anthropic's own admin console rather than the Office marketplace, is limited to Claude Team and Enterprise plans.

Microsoft's bet is that the platform layer abstracts the model layer. Anthropic's bet is that frontier labs do not have to accept being routed to. Both can be right at once, and they imply very different market structures three years out. The "which AI provider do we bet on" question is answered: you do not. Any strategy built around a single named vendor is vulnerable. Design around tasks, workflows, and data governance. Microsoft's public admission that no single model is reliable enough is also an architectural endorsement of verification-by-design. Any firm workflow that outputs AI-drafted content without a second-model check is the odd one out. The legal AI vendor moat has narrowed with it. Claude for Word meets lawyers in Word (like the Legal AI plugins), with a frontier model behind it, for the price of a Claude Enterprise seat. Vendors with deep workflow orchestration, proprietary data, and integration plumbing will survive the compression. The ones selling a thin wrapper on a frontier model will feel this first. Procurement teams that have been avoiding the "which AI stack are we building on" conversation are about to have it forced on them.

Claude Mythos Preview is a new frontier model with cybersecurity capabilities that surprised even Anthropic's own researchers. During internal testing, Mythos autonomously discovered thousands of zero-day vulnerabilities across every major operating system and browser, including a 17-year-old remote code execution flaw in FreeBSD that the entire security industry had missed. Rather than shipping commercially, Anthropic launched Project Glasswing: a restricted programme limiting access to around 50 partner organisations including Amazon, Apple, Microsoft, CrowdStrike, and the Linux Foundation, backed by $100M in usage credits and $4M to open-source security organisations. The model scores 93.9% on SWE-bench Verified, higher than any publicly available model, and it is not being sold. It is being deployed as defensive infrastructure. Sceptics counter that zero-day discovery is already industrialised through fuzzing, that severity matters more than raw volume, and that a capability-so-dangerous-we-cannot-sell-it narrative is convenient positioning in Anthropic's pre-IPO window. Both reads can hold at once.

Whatever is frontier today is commodity within 12-24 months. If Anthropic has built a model that finds vulnerabilities humans missed for 17 years, others are building it too, and not all of them will withhold. For law firms, which remain high-value targets (Jones Day was hit by Silent Ransom Group for a $13M ransom this same week), the attack surface just changed. Any cybersecurity posture that assumes human-speed offensive tooling is already behind. The firms investing in AI-augmented security testing now are buying time. Anthropic's decision to withhold the model from the market sets a separate precedent. The Bank of England has already added Mythos to its next Cross Market Operational Resilience Group agenda with Treasury, FCA, and NCSC attending, the first UK regulator to treat a frontier model as an operational-resilience question rather than a compliance one. For any innovation function that still thinks of cybersecurity as an IT problem, the Mythos moment closes that gap. Invest in the capability and invest in the governance to use it safely. Do not assume the next frontier lab will choose responsibility over revenue.