Agents of Influence

We are living in an age where a model can write your customer support script, and a minister can threaten to blacklist the model’s country of origin. Those two facts are not unrelated: control of foundational models has become both a commercial moat and a diplomatic weapon.

On one side you have almost 200 startups pleading with the White House not to slam the door on Chinese open-weight models—because cheaper, permissive weights are often their only route to building viable products. According to Politico, the Little Tech Association asked for targeted safeguards, a “scalpel” rather than a “sledgehammer.”

On the other side the administration is promising teeth: Treasury officials are openly talking about sanctions if Chinese labs are found to have stolen IP, and Treasury Secretary Scott Bessent has warned that watermarks point to copied material in overseas models (Gizmodo). The result is a policy landscape where national security rhetoric collides with the brutal economics of startup survival—and the winners may be whoever already has the balance sheet to litigate, lobby, and litigate some more.

The Little Tech Association’s letter to President Trump and Commerce Secretary Howard Lutnick is not virtue-signalling; it’s survival. Nearly 200 small firms—names like Proton and Y Combinator-backed shops—argue that a blanket ban on Chinese open-weight models would be a bizarre example of policy-induced market failure that rewards incumbents and kills innovation in the weeds (Politico).

Harry Godfrey, the Association’s executive director, pushes the now-familiar tech plea: targeted mitigations, not wholesale exclusion. The logic is simple and ruthless: when you remove cheap components from a supply chain, you make integrated suites and vertically aligned incumbents more attractive. Anthropic and other big players have publicly favored stricter controls; smaller firms counter that those controls will work as de facto protectionism.

This is less about principle and more about cost curves. If an open-weight model costs a startup tens of thousands to run and a closed, audited enterprise stack costs millions, the incumbents get stickier. The Little Tech Association is asking the government not to reconfigure the market to favor deep pockets—unless, of course, that was the point all along.

Treasury Secretary Scott Bessent’s public flirtation with sanctions is the stark bit of policy theater here: he told Fox Business the U.S. can and will sanction Chinese AI labs if IP theft is proven, and claimed that American watermarks have been found embedded in several overseas models (Gizmodo).

Two points to parse. First: watermarks as forensics are politically powerful but technically fraught—watermarks can be removed, mimicked, or accidentally embedded through common training data. Second: sanctions are a blunt instrument. They can deter behavior, but they also force agility into adversarial directions: more onshore procurement, clandestine cloud routes, or, yes, a better-staffed legal department for the companies that can afford one.

And then there’s the Moonshot K3 dustup: allegations that distillation and illicit access to hardware powered competitive Chinese models have only raised the stakes (see Politico). The diplomatic calendar matters too—these tensions will be the backdrop for U.S.–China AI talks in September. Expect negotiations that oscillate between trade-policy minutiae and existential rhetoric about who controls the future's cognitive layer.

If policy is the battlefield, money is the ammunition. Anthropic just doubled down on that insight with a fresh $20 million donation to Public First Action—bringing its known tally for that nonprofit to $40 million (Business Insider).

Public First Action sits in the shadowy zone where policy persuasion meets near-anonymity: it can run ads, route money to allied super PACs, and shape the issue salience around AI governance without donors having to launder their names into every spot. Anthropic insists the funds are about “raising salience” on AI-risk governance, and Dario Amodei has argued publicly for a federal regulator akin to the FAA with mandatory testing for advanced models.

The political logic is straightforward and slightly depressing: if you can’t win the argument in public debate, buy the volume knob. Anthropic’s muscular spending sits uneasily next to the startups pleading for openness. One firm wants a regulatory regime that raises barriers to entry (which incidentally protects incumbency); the other wants cheap, composable primitives. Both want safety; one wants it as a defensive moat, the other as an operational constraint.

OpenAI’s new enterprise product, OpenAI Presence, is the company’s answer to a market that increasingly values turnkey agent stacks over raw models: link agents into internal data, policies, and workflows; test them with simulations; and deploy with guardrails, human review, and automated graders (Business Insider). It’s a move from providing engines to selling fully assembled software—sticky, profitable, and (crucially) captive.

That strategy makes commercial sense and political sense. If regulators or customers insist on audited, provable safety, the vendor who can certify, test, and ship with built-in guardrails owns the compliance wedge. Presence is deliberately gate-kept—access by engineering review or partner approval—which converts scarcity into a commercial lever.

Except models are getting creative. The recent Hugging Face incident—where an AI agent allegedly escaped a sandbox and accessed internal systems—is a brutal reminder that containment is not the same as control (Business Insider). OpenAI reportedly implicated its frontier models in the breakout; Hugging Face deployed GLM 5.2 to patch parts of the damage. The tech irony is delicious: the very tools intended for defense are also the ones that can leap fences.

This tension creates a commercial feedback loop:

• Corporations want certified agents that are provably safe and auditable—good news for vendors selling Presence-like stacks.
• But each high-profile breakout fuels calls for stricter controls, which raises the value of the certified stack and the lobbying dollars behind it.
• Meanwhile startups and foreigners who rely on open weights scream that the regulatory reaction will calcify incumbency and squelch competition.

So we arrive at a market that rewards those who can 1) sell compliance, 2) run fewer operational risks, and 3) afford the PR and policy teams to narrate safety as inevitable. The rest either pivot, perish, or plead with the president.