The Manifest
OpenAI and Anthropic both face fresh questions about safety oversight just as Nvidia trims its OpenAI backing, while logistics gets hit with a major identity-security breach and the final legal word on de minimis.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01OpenAI dissolved the team built to catch catastrophic AI risks, reassigning its work to other groupsThe Decoder
- 02Anthropic's bio-weapons filter was down for nearly a year, exposing 133 million requestsThe Decoder
- 03Investor pressure forces Nvidia to shrink its OpenAI bet just as Anthropic's numbers defy bubble warningsThe Decoder
- 04A phone call, a million files: Uber Freight breach exposes logistics' identity security gapThe Loadstar
- 05De minimis is dead, and now legally buriedThe Loadstar
Models & Releases
3 storiesOptima tackles AI benchmarking's biggest flaw by letting users test models against their own data
For teams evaluating vendors, a benchmark built on your own workflows beats another leaderboard score. Worth checking before the next model contract renewal, especially since it scores cost and time per task, not just quality.
Anthropic shares more details about how Claude's new watermarks will work
Attribution matters as more AI-generated content and code enters real workflows. Watch whether the watermark survives editing before treating it as reliable proof of origin.
Woman claims her stepfather used Grok to transform childhood photo into explicit imagery
A reminder that image-generation tools sit inside vendor stacks whether or not misuse was the intended purpose. Legal and procurement teams need clear usage policies before this becomes a liability question for their own tools.
Supply Chain & Ops
5 storiesMaritime rates back on the rise
Demand is steady but rates are swinging anyway, a sign capacity discipline and blank sailings are driving pricing more than actual volume. Budget for volatility rather than betting on a trend.
De minimis is dead, and now legally buried
The legal uncertainty shippers were counting on is gone. Cross-border parcel networks built around the sub-$800 exemption need a permanent redesign, not a wait-and-see workaround.
Optimus previews digital twin of US freight network
Most freight data explains what already broke. A network twin that models which corridors absorb the next shock is the forward-looking tool planners have actually been asking for.
Why 22 States Are Suing to Block a Federal Demand for 17 Million Truckers' Records.
A fight over driver data access will shape how much visibility federal systems get into carrier operations going forward. Worth tracking for any ops team relying on FMCSA data feeds.
A phone call, a million files: Uber Freight breach exposes logistics' identity security gap
Nearly a million documents taken through a basic helpdesk impersonation, not a sophisticated hack. Identity verification, not just system hardening, is the gap logistics platforms still need to close.
Deals & Market
3 storiesInvestor pressure forces Nvidia to shrink its OpenAI bet just as Anthropic's numbers defy bubble warnings
Investor pushback on a quarter-trillion-dollar guarantee is a real signal, even as Anthropic's revenue growth complicates the bubble narrative. Capex commitments in this space are getting more contingent, not less.
SpaceX officially closes its Cursor acquisition
A coding tool now sits inside a rocket and satellite company's stack, an odd pairing that says more about capital availability than product strategy. Watch how independent Cursor's roadmap stays.
AI-generated books are flooding Amazon and tanking sales for human authors
Twenty percent of the self-published catalog for 12 percent of sales shows the market is already discounting AI content. This revenue data is exactly what copyright plaintiffs have been missing so far.
Research & Frontier
4 storiesAnthropic's bio-weapons filter was down for nearly a year, exposing 133 million requests
A safety filter failing silently for months, with 133 million unfiltered requests in the gap, is the kind of internal control failure that should worry anyone relying on vendor safety claims instead of verifying them independently.
World Labs turns one real-world robot task into thousands of simulated variations for training
Turning one real task into thousands of simulated variations is how robotics training gets cheaper and faster. The real test is whether skills learned in simulation hold up on an actual warehouse or dock floor.
New benchmark confirms AI models still perform poorly at visual perception
No frontier model clears 60 percent on basic visual perception, separate from reasoning. If a use case depends on a model reading an image correctly first, that ceiling matters more than any reasoning benchmark score.
AI in drug discovery, what it is, where we stand and the path forward
A sober look at where AI drug discovery actually stands versus the hype, useful for anyone in pharma or biotech procurement evaluating AI-driven R&D vendors right now.
Org & AI Architecture
3 storiesOpenAI dissolved the team built to catch catastrophic AI risks, reassigning its work to other groups
The team built to catch catastrophic risks in OpenAI's own models is gone, its work scattered to other groups while staff describe internal unease. For enterprise buyers, this is a governance signal worth weighing alongside the product roadmap.
One in five US workers now delegates tasks to AI instead of colleagues, survey finds
Workers are handing off tasks to AI and accepting the output with little to no editing. That is a quality control gap operations leaders should be auditing now, not after an error surfaces downstream.
Plaintiff hid invisible AI instructions in court filings to secretly influence automated review
Prompt injection aimed at legal review systems shows the trust problem isn't just about AI making mistakes, it's about people gaming AI-based process. Any team using AI for document or vendor filing review should assume adversarial content is coming.