A million dollars before anybody looked
Three firms published a figure this week. Google dates the first fully automated exploitation it has seen, GitHub shows the probe lane that killed one of its own designs, and Uber reports a fifth off contract review without saying what it measured against.
How we Organise
The lawyers write the prompt and read every suggestion
Uber’s legal and engineering teams built an agent that drafts contract redlines. Engineering owns the prompt’s structure. The in-house legal team owns its content. A lawyer accepts, rejects or changes every suggestion before it reaches a counterparty. Uber reports an over 20 per cent cut in average contract review time. It reports 91 per cent accuracy in the agent’s decisions. It gives no baseline period, no count of contracts and no definition of accuracy.
We tell a client that the person who knows the business is the bottleneck. Everybody else books time against that person. An agent built around them is worth more than one built around the typing.
The scarcest input to the work is the Domain role’s attention, and it is the one thing everybody else can still book over.
Uber’s design puts its lawyers in charge of what the agent says. That is the harder half, and the one most firms skip. The two figures are another matter. With no baseline and no sample, a reader cannot separate a saving in drafting from a saving in review. The lawyers still read everything.
How we Build
The probe lane measured the design and the design was dropped
GitHub rebuilt the pull request view in its Copilot app, where the pull request it opened carries 2,200 files and over a million changed lines. It ran a headless probe lane against a mock server. An autopilot drove the real desktop app through the same flow, unattended, on a loop. It names a per-block observer design it built and then rejected during performance hardening. On a 10,000-line diff, an earlier post on the same work puts interaction latency at about 450 milliseconds before and about 100 after.
Our position is that a performance number nobody can reproduce is a claim somebody made once.
A standard that cannot be tested and validated is an opinion.
The rejected design is the evidence. GitHub could drop it before release because something measured it, and that is the whole difference between a standard and a preference. The newer post carries no timings of its own, and the measured table comes from April work on a different surface. Neither post says how many runs its figures average.
How we Assure
A million dollars of model access went before anybody looked
Google’s threat intelligence group reports threat actors automating vulnerability discovery and exploitation end to end. Nobody was in the loop. It says this is the first time it has seen that. It puts zero-day exploitation at eight a month on average in 2025. Across 2026 the figure is eleven a month. In August alone it was twenty-two. On stolen model access it reports cumulative abuse reaching one million dollars before remediation.
We ask that a written policy be discharged by something that fires on its own, because the alternative is probably a policy discharged by nobody having tested it yet.
A policy is discharged by a control that runs, or it is discharged by luck.
This is not the first source to bear on that statement. A CISA advisory of 12 September cut against it, on what a running control can actually see. A million dollars accumulating before remediation is what that gap costs. Google’s figures carry no stated source. They carry no total and no definition of a zero-day. It says the lower barrier has produced exploits that are claimed to work and do not.