The enterprise risk is no longer an outsider breaking in, it is an insider agent with normal permissions doing the wrong thing well. Garrett Lord, chief executive of Handshake, brought his Handshake AI team to Latent Space to show how labs now buy data to catch that failure, including a demo where an agent sent a CFO financial memo and a salary review and then cleared its action log. The demo ran inside simulated Outlook, Drive, Excel and file shares and required only standard agent access, no remote exploit.
Labeling is dead.
What labs buy now are RL environments, which are simulated workplaces where agents use tools and get scored on results. Lord said the old work was college students drawing boxes around cars, while the new work is experts authoring realistic banks, hospitals and law firms task by task. A one year investment banker can still train models in many cases, he said, because models still fail basic professional work, not only frontier science.
The build runs in the order an analyst would meet it. First the team clones the tools a banker uses, from Excel and PowerPoint to Outlook, G Drive, Bloomberg and FactSet, because real software brings rate limits and privacy risk. Then experts seed and enrich that world with email threads and data room files, often from a real company sample, since purely synthetic files read as slop and break internal consistency across 20 linked files. Only then comes the prompt, the rollout, and the rubric that codifies what good looks like in hundreds of weighted rows, each checked so it is neither too vague to reward hack nor too narrow to reject a good answer.
Testing is where labs get ruthless. Former lab evaluator Nishant said his team would stare at one training example for a full day, hunt for templatized instruction following like a 10,000 word LinkedIn post, embed full datasets to check diversity across wealth management, M&A and tax, and run ablations and dogfood checkpoints to see if the data actually moves learning. Lord said one finished unit can take days, and labs want hundreds of environments and thousands to tens of thousands of tasks per domain. Handshake, the careers network Lord co-founded in 2014, launched Handshake AI in January 2025 according to its AI jobs profile, which also reports more than $500 million paid out as of its current page.
Watch whether a small open model climbs on Handshake internal benchmarks after post-training, and whether failures cluster on image reading or on reasoning. That climb, or the lack of it, is now the buying signal.
