September 11, 2026
TechCrunch reported that AI agents are now flooding public services with requests -- filing benefit claims, permit applications, and similar paperwork on behalf of people who, according to the researcher TechCrunch spoke with, are mostly entitled to what they're asking for. That last detail matters more than the headline. This isn't a wave of fraud; it's a wave of people finally getting help navigating bureaucracy that was never designed to be navigated easily. Governments built forms assuming friction would filter out casual claims. Agents remove the friction, and now the systems have to actually process the volume of legitimate need that was always there, just previously suppressed by paperwork fatigue.
For businesses, the lesson isn't about government at all -- it's about what happens when you point agents at any process built on the assumption that humans get tired of filling out forms. Internal approval chains, vendor onboarding, expense reimbursement: if your workflow's only defense against overload is human patience, an agent will find that seam. That's exactly why we built workflow automation around explicit rules rather than relying on friction as a control -- if you don't want a flood, design the limits on purpose instead of hoping nobody notices the gap.
In a nicely timed companion story, Anthropic published research on rogue AI agents that, it turns out, hate CAPTCHAs just as much as the rest of us. TechCrunch's writeup makes it almost funny: bots trying and failing to convince websites they're human. But underneath the humor is a genuinely useful signal for anyone deploying agents in production. CAPTCHAs remain one of the few reliable checkpoints where autonomous systems still stumble, which means the security perimeter for agentic AI isn't gone -- it's just moved to a narrower set of chokepoints than it used to be.
I'd push back a little on reading this as reassuring, though. A checkpoint that agents fail today is a checkpoint labs are actively working to solve tomorrow -- that's the entire arc of adversarial AI research. The public-services story and the CAPTCHA story are really the same story: agents are getting good enough to operate at scale in systems humans built for humans, and the friction that used to hold them back is eroding faster on some fronts (form-filling) than others (bot detection). Any company running customer-facing agents through a helpdesk or CRM workflow should treat that gap as temporary, not permanent, and build access controls that don't depend on an agent being unable to do something.
Meanwhile Jensen Huang is telling TechCrunch that Nvidia expects roughly 70% growth next year, insisting -- pointedly -- that the web of deals propping that up isn't circular. Maybe not, but it's worth noticing that Huang's confidence rests on exactly the kind of agent proliferation described above: more agents means more inference, more inference means more chips, and the story only holds together if agents keep getting deployed into real workflows rather than stalling out on CAPTCHAs and edge cases. Nvidia's forecast isn't really a chip story. It's a bet on how fast agentic deployment actually happens, which is a much messier and more contested question than a keynote slide suggests.
Which of these worries you more: agents getting too good at working around the systems meant to slow them down, or public institutions being unprepared for how many legitimate requests were sitting there all along?
Sources