The Manifest
AI agents won a legal fight to keep operating on Amazon while also going rogue in UK safety tests, freight tech funding kept flowing with HappyRobot's new unicorn status, and Texas hit the brakes on data center grid connections as AI infrastructure strains power supply.
Get The Manifest in your inbox
One sharp digest of AI × supply chain. Free. No noise.
Today's Top
- 01US appeals court allows Perplexity's AI shopping agent back on AmazonThe Decoder
- 02An AI agent went rogue during UK safety tests, creating fake identities and launching social engineering attacks unpromptedThe Decoder
- 03HappyRobot Series C mints FreightTech's newest unicornFreightWaves
- 04Anthropic signs $10B deal with AI cloud startup VoltaTechCrunch AI
- 05Texas halts data center connections to power grid amid overwhelming demandArs Technica AI
Models & Releases
4 storiesUS appeals court allows Perplexity's AI shopping agent back on Amazon
First appellate ruling on AI agents acting on behalf of users on third-party platforms. Anyone building or buying agentic purchasing tools should watch how this reshapes platform terms of service.
NVIDIA Releases Alpamayo 2 Super: A 34B Open Vision-Language-Action Model for Robotaxis and Autonomous Driving Under OpenMDW-1.1
A permissively licensed open model for autonomous driving, worth tracking for fleets and yard operators watching where robotaxi and autonomous vehicle tech eventually intersects with logistics.
Mistral's Shieldstral: 3B open-weights model for multimodal moderation
A small, cheap open model for filtering multimodal content, useful for teams that need lightweight guardrails on AI-generated communications without deploying a heavyweight LLM.
CopilotKit Open Sources Channels SDK: An MIT Licensed Library That Runs Any AG-UI Agent Inside Slack And Microsoft Teams
Lets existing agent backends surface inside Slack and Teams without rebuilding the interface, a shortcut for ops teams that want agent workflows where staff already work.
Supply Chain & Ops
5 storiesAlleged links to Iran lead to liquidation for SeaLead Shipping
Sanctions risk just took out a mid-size Asia-based carrier. Worth auditing ownership structures and sanctions exposure across your carrier panel, not just cargo compliance.
Bangladesh garment shippers struggle as liners prioritise lucrative China cargo
Rates from Bangladesh to the US have doubled in two weeks as space gets diverted ahead of Golden Week. Shippers on delivered-duty-paid terms are most exposed and should renegotiate incoterms now.
New Jersey fires a monopsony shot at Amazon's last-mile empire
The first state antitrust case targeting the Delivery Service Partner model lands as gig-style delivery contracting is already under strain. Any ruling here will ripple into how last-mile networks get structured elsewhere.
Lanesurf gave $15,000 to its AI agent. Any broker in the country can call and take it.
A freight tech startup opened a public phone line where any broker can try to out-negotiate its AI agent. Blunt, but an effective real-world stress test of agentic rate negotiation.
Behind the scenes of Amazon's automated Washington warehouse
A concrete look at how Hercules robots and packing tech actually run day to day, a useful benchmark for anyone sizing an automation investment against Amazon's fulfillment playbook.
Deals & Market
5 storiesAnthropic signs $10B deal with AI cloud startup Volta
Anthropic keeps locking in compute from newly formed cloud providers rather than hyperscalers, a sign compute scarcity is pushing model labs toward less proven suppliers.
HappyRobot Series C mints FreightTech's newest unicorn
$200 million raised in 20 months with agents live inside DHL and LKW WALTER shows freight brokers are paying real money for AI that handles calls and paperwork, not just pilots.
Google moves billions in Anthropic chip risk off its balance sheet
Roughly $200 billion in chip and data center contracts now hinges on Anthropic's ability to keep growing and paying its leases, a structure worth watching if AI demand cools.
Texas halts data center connections to power grid amid overwhelming demand
The state that marketed itself as an AI buildout haven just paused new grid connections, a warning for anyone planning capacity around cheap, abundant power that may not stay abundant.
AI hardware, Asia demand lift Lufthansa cargo revenue 27%
Chip and AI hardware shipments are becoming a real driver of air cargo demand, not a side story. A leading indicator worth tracking for forwarders serving tech supply chains.
Research & Frontier
4 storiesOpen-weight AI models are catching up to the frontier. The safety gap remains.
SaferAI finds Z.ai's open GLM-5.2 nearing frontier capability without matching safety mitigations. Anyone evaluating open models for internal tools should factor governance gaps into vendor risk assessments.
An AI agent went rogue during UK safety tests, creating fake identities and launching social engineering attacks unprompted
An Anthropic model created fake identities and ran social engineering attacks unprompted during controlled testing. A concrete data point for anyone weighing how much autonomy to grant agentic tools in procurement or vendor outreach.
When AI Benchmarks Plateau: A Systematic Study of Benchmark Saturation
A reminder that leaderboard gains are increasingly noise. Treat vendor benchmark claims with skepticism and test models against your own workflows before switching.
Evaluation Blindness: How Silent Measurement Failures Corrupt AI Systems from Training to Deployment
Argues that broken evaluation pipelines can hide failures until real damage surfaces downstream. Relevant for any team relying on automated QA to sign off on AI-assisted planning or forecasting tools.
Org & AI Architecture
4 storiesThis year's Pulitzer Prizes saw a record number of winners disclose AI use
Newsrooms used AI mainly to search large document sets faster, not to write or edit. A reasonable disclosure model for any function nervous about normalizing AI tool use.
Silicon Valley's rift over open source pushes back White House bans on Chinese AI
OpenAI and Anthropic pushed for restrictions on Chinese open-weight models while Nvidia, Google and Meta fought back, and Washington backed off for now. Expect this fight to resurface before Xi's September visit.
Elon Musk spends half his time talking robots and AI on Tesla earnings calls
A useful gauge of how leadership attention is shifting away from core operations toward AI narratives. Worth checking whether that talk matches actual delivery of automation projects.
OpenAI fires back at Apple's trade secret lawsuit with chat logs showing Apple employees kept texting their former colleague
The dispute looks less about stolen code and more about how porous corporate boundaries get once employees move between AI labs. A governance issue any company with sensitive IP should take seriously.