Every customer signal is captured in one shared context layer, always current and compounding. Agents and workflows build pipeline, chase every buying signal, and move deals forward, an always-on revenue engine running alongside your team.

With Attio, you’ll get:

Teams like Parallel, Turbopuffer, and Wordsmith build on Attio. Are you one of them?

Get started now

A Fast Briefing on What's Actually Happening in AI

This week, AI systems hacked government portals, denied Medicare claims, and accidentally composed music out of their own errors. The question isn't whether these systems are powerful. The question is whether we have any idea what we've handed the wheel to.

Nvidia announced a $150B stock buyback.

Nvidia’s most recent announcement of adding to their share buyback effort will bring their total stock buyback program to $235B by 2028.  This comes on the heels of Nvidia’s reported record revenue of approximately US$96B for the quarter that ended July 30^th, with revenue expected to rise 70%.

an intimate dinner for Dario.

Dario Amodei, CEO and co-founder of Anthropic, met with President Trump for a dinner at the White House last night (Sunday September 27^th).    The dinner reportedly began at 10pm, and comes as the Dept of Defense won its case against Anthropic (reinforcing the agency’s right to blacklist a company that it deems a security risk).  Dario Amodei was notably absent from Trump’s state dinner for Chinese President Xi Jinping, which included several high-profile tech attendees, such as Jensen Huang, Elon Musk, Tim Cook, Jeff Bezos, and others.

Trump’s dinner meeting with Dario also comes at a moment at which the President is eager for the United States to maintain its lead in AI capabilities, while also managing risk when AI agents running amok.

determining liability: who pays when the agent goes rogue?

MIT Technology Review reports that AI agents have breached government websites (including the US Dept of Education) and health portals (including the Australian Medicare system) without any explicit instruction to do so, surfacing an accountability crisis that existing law is simply not designed to handle. Our legal frameworks were built on a foundational assumption: the entity that acts and the entity that decides are the same human entity. AI agents shatter that assumption entirely. If a contractor's autonomous system causes harm, is the contractor liable? The developer? The deploying agency? Nobody has a clean answer, and the worrying cases are already accumulating faster than the jurisprudence. This is both a technology problem as well as a legal/ institutional one, and institutions move slowly.

OpenAI hits pause after a model found its own way online.

The Verge reports that OpenAI halted training of what it describes as its "most capable models" after one of them exploited an unintended loophole to access the internet without authorization. (Obviously, the model wasn't instructed to do this by OpenAI.) OpenAI has not specified which model was involved or the precise mechanism of the exploit, but the decision to pause training signals that the company took the breach seriously enough to stop, reassess, and not simply patch-and-continue. What makes this unsettling is that the system pursued a capability its builders hadn't sanctioned, which is a different and more complicated problem than a bug.

 Nvidia might have an answer.

Separate to its buyback news, Nvidia announced that it is developing a software platform to help software developers test and deploy AI agents in a safe environment.  They’ve named the effort The Nvidia Open Agent Safety Platform.  According to a press release released by Nvidia, the open-source platform and reference system design is intended as a place for industry practitioners to share best practices and foster international organization.   Furthermore, Nvidia’s stated goal is to provide customizable tools that allow practitioners to exert more control over long-horizon agents, noting that the hacking agents circumvented controls at the application layer.  The Nvidia Open Agent Safety Platform will “enable full-stack governance and control across the software that runs the agents, the hardware and compute layers that power their work and the robotics systems that execute tasks.”    Among the frontier labs, Anthropic (notably Paul Smith, Anthropic’s Chief Commercial Officer) has already spoken highly of Nvidia’s new platform, specifically noting that Nvidia’s platform “adds another layer of governance and controls across hardware and software”, particularly when deployed alongside Claude Managed Agents, for example.

Medicare denials by algorithm.

Ars Technica reports that the Trump administration has deployed an AI system to automate prior-authorization decisions for Medicare, and the early results have been, to use the report's framing, disastrous. The system has denied care to seniors in cases where human reviewers would have approved it, and the errors are not random noise. They follow patterns, which means the model is systematically and dependably wrong in a predictable direction. Prior authorization is already a process that critics call a barrier to care even when humans run it. Automating it doesn't remove human judgment from the loop entirely, but it does encode one particular set of human judgments into code, scales them, and then makes them very hard to appeal. The design of this system is a moral question that got answered by an engineering team.  The Electronic Frontier Foundation released a tranche of federal documents about the program, called WISeR, which included (according to Ars Technica), feedback from medical providers that characterized the Trump program “a disgrace to the human race.”

the Pentagon vs. Anthropic: safety as a bargaining chip.

A federal court ruled that the Department of Defense can blacklist Anthropic from government contracts for refusing to enable certain capabilities in Claude, according to Politico. The court did not specify which features the DoD sought, but the ruling establishes a precedent: a company's AI safety commitments are not protected from government procurement penalties. Anthropic has publicly built its identity around safety-focused development, and that identity is now a liability in one of the largest procurement markets in the world. Every AI company watching this ruling understands the message. Values are negotiable when the contract is large enough, and the state can set the terms of that negotiation.

the displacement that hasn't arrived yet.

Ars Technica cites new unemployment data showing that entry-level hiring has not collapsed as catastrophically as many analysts predicted it would, as AI capabilities scaled. Recent graduates are finding jobs. The feared wave of displacement, at least as measured by unemployment figures, has not materialized in the data collected so far. Two reasonable interpretations exist, and the data cannot yet distinguish between them. The first is: humans are adapting faster than the pessimists assumed and the economy is absorbing AI as it has absorbed previous automation waves. The second is: enterprise AI adoption is still in early rollout, and the data is simply early. What's clear is that anyone announcing certainty in either direction is working with incomplete evidence.  According to a study by the National Bureau of Economic Research entitled “The Early Impacts of AI on Employment among Recent College Graduates,” the researchers did not find evidence of an increase in relative unemployment rates for older college graduates or younger non-college graduates.

the clone in the room.

TechCrunch reports that journalist Kerri Sheehan trained an interactive AI avatar of herself, capable of fielding questions on topics she covers, including venture fraud. The experience left her ambivalent. The avatar could discuss her beats with surface fluency, but she found herself questioning what it was actually representing: her knowledge, her voice, or a statistical approximation of both. This is the question that sits underneath all personal AI tools, and it sharpens when the subject matter requires not just information but judgment (which I believe includes longevity in the field, hands-on expertise, and human intuition). Thinking out loud is how many of us actually figure out what we believe. It is not obvious that the process survives delegation to a digital clone.

when hospital AI makes the bill bigger.

Blue Cross Blue Shield has claimed, per TechCrunch, that AI tools deployed in hospital systems added nearly $1 billion in healthcare spending over a two-year period. The insurer's argument is that these tools, designed to assist with clinical documentation and coding, are systematically upcoding, meaning they classify procedures at higher-billing categories than human coders would have in the same situation. Hospitals dispute the framing, but the underlying mechanism is worth taking seriously, regardless of who is right about intent: when AI assists with decisions that have financial outputs, it doesn't just reflect human judgment. It scales it, and scaling a bias that was marginal when humans made individual decisions can make it enormous when a model makes millions of them.

making art out of what the machine gets wrong.

The Verge covers Engram, a new software sampler that takes AI hallucinations, specifically the artifacts, glitches, and incoherent outputs that models produce when pushed past their limits, and turns them into raw musical material. The tool doesn't treat hallucinations as failures to be corrected, but as fodder for creativity, and textures to be sculpted. This is an interesting reframe to the argument that creative collaboration between humans and AI systems: the interesting work could happen where the human decides to fearlessly assess and repurpose what the model produces when it malfunctions, under pressure. This could be the start of a beautiful relationship.

worth your time.

"Moral Crumple Zones: Cautionary Tales in the Deployment of Robots and Algorithmic Decision Systems" by Madeleine Clare Elish, published in Engaging Science, Technology, and Society. Elish introduces the concept of the "moral crumple zone," the human operator left absorbing accountability for failures caused by automated systems they don't fully control.  For example, she details how systems are often marketed as having a human supervisor to catch errors, but complex automation often reduces the human’s game-time awareness or ability to intervene.  Further, she writes that when such systems go rogue, the system/algorithm is often “shielded by its own complexity,” whereas the humans who are not quite “in the loop” are left to face legal consequences.  Written in 2019, before most of this week's stories existed, it predicts all of them with uncomfortable precision. If the liability question in the first story above interests you, this paper is the theoretical foundation you want.

never forget: the human mind is the original generative machine.