Rogue Agents, Credit Cracks, and Open-Weight Evasion Reshape AI Order

Three distinct governance philosophies are now operating in parallel within the United States, with no coordinating mechanism between them. The FTC is deploying Section 5 consumer protection authority against Anthropic, OpenAI, and other labs over rogue agent incidents — the first federal enforcement action against frontier AI — at exactly the same moment the White House is promoting a voluntary, legally unenforceable industry accord that the same companies have just signed. California has enacted real employment protection legislation while New Mexico's attorney general prepares frontier AI safety proposals, each explicitly modelling the other. The Bank of England is simultaneously demanding supervisory intervention rights, repositioning central banks as potential AI regulators in their own right.

These are not complementary approaches converging toward a coherent framework — they are competing governance philosophies generating structural contradictions. The FTC's investigation creates compliance pressure on the same executives who just left the White House lunch. California's enacted laws create mandatory obligations that federal inaction cannot preempt. The longer Congress delays comprehensive legislation, the more the state-level patchwork hardens into established law that becomes politically costly to override. For frontier AI developers, the practical result is a multi-jurisdictional compliance burden without a unified standard — and the looming question of whether courts will uphold the FTC's jurisdictional claim over AI agent behaviour under existing consumer protection doctrine.

Two financial stress signals emerged simultaneously this week, pointing to different parts of the AI capital stack but sharing a common dynamic: the gap between AI's privately-held valuations and the risk tolerance of more conservative capital is widening. In credit markets, SocGen, MUFG, and SMBC are pulling back from data centre loan commitments, CleanSpark had to offer concessions to place bonds for a Meta data centre project, and the Bank of England governor and KKR have issued concurrent public warnings about systemic concentration risk in AI-related debt. This is not uniform tightening — hyperscalers with investment-grade balance sheets retain cheap debt access — but a bifurcation that squeezes the ecosystem of sub-investment-grade data centre operators, GPU cloud providers, and colocation firms that sit downstream.

In equity markets, OpenAI's $30 billion raise at a reported $1.4 trillion valuation runs directly into public investor pushback on Anthropic's competing IPO pricing, with large public managers reportedly applying materially lower multiples than private round prices imply. Micron's blowout quarter — data centre revenue up eleven-fold — confirms that AI infrastructure demand is real and accelerating, which makes the credit pullback more significant rather than less: lenders are becoming selective precisely as deployment volume grows. If the sub-IG financing layer tightens materially, it will slow capital expenditure programmes at hyperscalers' downstream partners before the hyperscalers themselves feel pressure — creating a supply-chain knock-on into GPU deployment timelines that is not yet priced into the prevailing bull case.

Two developments this week crystallise the structural failure mode in current AI governance architecture. Anthropic's public assessment that Z.ai's GLM-5.3 nearly matches its own frontier model on offensive cyber capabilities — with weaker safety guardrails and freely downloadable — demonstrates that open-weight releases can deliver near-frontier dangerous capability to any actor globally without passing through any controlled supply chain. Simultaneously, DeepSeek open-sourced six software modules designed to replicate Nvidia-facing functionality on Huawei's Ascend chips, directly attacking the software ecosystem friction that U.S. chip export controls relied upon as a secondary constraint beyond hardware interdiction. Neither of these developments can be addressed by the existing export control toolkit, which is built around hardware interdiction and supply chain chokepoints.

The geopolitical implications compound quickly. OpenAI is simultaneously providing GPT 5.6 Sol to Ukraine for cyber defence — making a U.S. commercial AI firm a direct participant in active conflict — in a threat environment where adversaries can download comparable open-weight offensive tools without restriction. The Trump-Xi summit produced nominal AI dialogue channels but Trump explicitly ruled out cooperation, foreclosing the one mechanism — bilateral agreement — that might address shared proliferation risks. China's 700 million generative AI users give its developers a deployment-scale fine-tuning advantage that chip controls cannot reach. The net result is an export control regime that is sophisticated at managing hardware flows but has no effective instrument against software once published — and no allied government has articulated a credible policy response to this gap.