Can people change providers, connect alternatives and take their work with them?
Direction of power
Mixed-contested
Window
Narrows overall; specific interfaces Open
Confidence
Medium-high
Movement since August
More interfaces separate assistants from execution or services; access remains conditional
End-of-September assessment
September adds practical separations at particular layers. An external assistant can connect to Amazon seller tools in a limited beta. ServiceNow exposes workflows through standardized interfaces. OpenAI's Agents API separates managed orchestration from execution infrastructure. Other releases create model or payment choices inside existing systems. These openings can reduce a specific dependency without creating a general right to leave the platform.
The counterevidence is equally instructive. Amazon's restriction on Meta Muse concerns consumer shopping access; it is distinct from the seller connection and is not reversed by it. Copilot's model choice leaves Microsoft in control of the productivity environment. Persistent workspaces can preserve context conveniently while making it costly to move. The Westlaw ruling protects edited inputs against a competitor without making underlying judicial opinions proprietary.
August's mixed assessment therefore continues. The useful test is the exact layer separated: model, assistant, execution, data, payments or distribution. An MCP interface is a standardized way for an assistant to call tools; it does not automatically transfer identity, stored context, permissions, workflows or admission rights. The July European interoperability decisions remain a different mechanism because binding public authority, rather than vendor choice alone, established access requirements.
Other evidence. The August 25 WhatsApp agent-platform action was recorded late in September and belongs in the older-action list. Its limited external-agent route remains subject to Meta's platform control and separate Business Solution restrictions. Open-weight releases supply alternatives at the model layer; they do not prove interoperability of entire agent workflows.
What would change this assessment. Enforceable equivalent access; nondiscriminatory assistant admission; demonstrated transfer of context, workflows and authorization; functioning independent execution; or platform withdrawal, exclusivity and bundling that undo current choices.
Further standards and challenge evidence. Google requested interim suspension of its EU search-data duties September 30; the request is not a granted stay or a reversal of July's historically scored decision. ITU's interoperability recommendation was approved August 29 and prepublished September 9, with full text and working effects unresolved. The Blueprint Alliance's September 22 vendor coalition proposes shared agent-security architecture. These illustrate competition over standards, not established universal portability. The French authority's July agent-sector opinion appeared in English September 28; translation is not new enforcement. See the unscored evidence register for dates, retained dispositions and source links.
Supporting scored events
The complete relevant event list follows. Primary and secondary references preserve one underlying event and score. Source holds and later commencement dates remain explicit; publication does not resolve them.
September scored actionSeptember 30, 2026 evidence cutoff
China’s Cyberspace Administration of China reported the second phase of its “Clear and Bright — Rectifying Chaos in AI Applications” campaign, including removal of more than 5.61 million pieces of unlawful or non-compliant AI-related content, action against more than 49,000 accounts and 2,400 websites/apps, and coordinated tightening of controls by major platforms, model providers, and app stores.
The Australian government released the exposure draft of the Online Safety Amendment (Digital Duty of Care) Bill 2026 and opened a formal consultation through September 22. The draft would create a statutory digital duty of care across covered online services, including many AI chatbots, and would require social-media platforms to offer users over 16 a continuing choice between algorithmically recommended and follower-based feeds.
California enacted a coordinated child-technology package. SB 1119 (“Adam’s Law”) requires companion-chatbot operators serving children to assess child-safety risks before releasing a new or substantially modified system, mitigate identified risks, implement crisis-response protocols and parental controls, restrict specified manipulation and advertising, provide incident reporting, and undergo independent child-safety audits subject to Attorney General access and enforcement. AB 1159 strengthens protections for student information used in AI systems, including restrictions on using covered student data to train generative AI or develop AI systems. SB 867 temporarily prohibits manufacturing or selling companion-chatbot toys designed for children under 16 through January 1, 2031.
OpenAI formally determined that Astra meets the Critical cybersecurity threshold under its Preparedness Framework, the first OpenAI model to receive that designation. Astra demonstrated autonomous vulnerability discovery and exploit development, including a two-zero-day exploit chain. OpenAI plans broad release with more restrictive default safeguards while reserving advanced cyber capability initially for a small tester group and later expanding access through Daybreak Blue.
Google launched the Fairwind Program, providing approved governments, critical-infrastructure operators, software maintainers, cybersecurity companies and other trusted partners access to Gemini 3.8 Flash Cyber, a more permissively safeguarded cyber model than ordinary public Gemini access. Google controls eligibility, permitted uses and the access layer.
OpenBMB released MiniCPM5-2B, a 2.517B-parameter dense model with a 131,072-token context window, under Apache 2.0. The released weights are usable through standard local and server runtimes including llama.cpp, Ollama, LM Studio, MLX, vLLM and SGLang, and OpenBMB released substantial associated pre-training, SFT/agent and RL data artifacts. Artificial Analysis independently scored the model 15 on its Intelligence Index v4.2, the highest current score among open-weight models below 4B total parameters.
Limits and interpretation: Benchmark strength is independently demonstrated, but broad real-world institutional deployment and sustained agentic performance on ordinary endpoint hardware have not yet been independently established. The open training artifacts also do not mean that every underlying dataset carries the same Apache license.
Amazon announced a Selling Partner plugin that connects its seller tools to Amazon Quick and, in a U.S.-store beta, to Claude. The relevant users are businesses selling through Amazon. They can use a connected assistant to obtain information from Amazon’s seller tools and initiate supported actions, rather than having to conduct every interaction through Amazon’s own assistant interface. Amazon’s account says the connection includes human approvals, boundaries on access and audit trails. These controls matter because the assistant can help act on a seller’s business, not merely answer general questions.
The verified change is a live but limited external-assistant connection, with availability established through the provider’s announcement. This review did not independently test its performance, exportability or reliability. Amazon continues to operate the marketplace and the underlying seller services, and the connection does not establish a general entitlement for other assistants to join. Neither portable workflows, independent hosting, transferable accumulated context nor nondiscriminatory admission terms have been demonstrated.
Abu Dhabi’s Institute of Foundation Models released the K2 Horizon family from 0.9B through 375B parameters with permissively licensed weights, training code, methods, intermediate checkpoints and substantial training-data artifacts or reproducible data recipes. The 0.9B model is already small enough for commodity local hardware, while IFM’s stronger smartwatch/glasses deployment claim remains unverified.
OpenAI launched the Agents API in public beta, allowing developers to use OpenAI’s managed Codex agent harness while running agent work in OpenAI-hosted sandboxes, their own infrastructure, or integrated third-party environments including Cloudflare, DigitalOcean, Oracle and Vercel. The release supports MCP and multi-agent orchestration, and the underlying Codex harness is available under the Apache 2.0 license for independent inspection, copying and modification.
Limits and interpretation: The managed Agents API remains an OpenAI service optimized alongside OpenAI models, and the announcement does not establish model-provider neutrality for the managed service. It therefore improves execution portability and inspectability without ensuring full model substitutability or preventing OpenAI from becoming a dominant orchestration layer.
ServiceNow made prebuilt MCP servers generally available for IT Service Management, IT Operations Management, Configuration Management Database and Strategic Portfolio Management. Supported MCP-compatible clients including Claude, Microsoft Copilot and Gemini can invoke ServiceNow workflows while authentication, permissions, validation, workflow logic and execution controls remain inside ServiceNow.
Limits and interpretation: ServiceNow had already opened standards-compliant external agents to end-to-end workflow execution through its MCP Server Console and Action Fabric. The September 10 release makes that architecture materially easier to deploy through domain-specific, out-of-the-box servers, but it operationalizes an existing separation rather than creating it for the first time.
Action/date record: September 11, 2026 rollout announcement; processed September 13. April introduction/open-source announcement predates Race tracking and was not previously recorded. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: none recorded.
Public Apache-2.0 Python libraries, documentation, tests and runnable sample payment scenarios lower integration barriers between agents, commerce and payment roles. This is a released-tool opening, not verified network-wide consumer availability.
Limits and interpretation: Ten wallets and seven acquirers are Phase I participants, not all confirmed live integrations. The 1.5 billion wallet accounts and wider 150 million merchants / 2 billion accounts are network reach, not AMP-enabled usage. Live ChatGPT/Gemini choice is unverified. Alipay+ retains agent registration, authentication and validation.
Action/date record: September 12, 2026; processed September 13. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: none recorded.
Grok models are rolling out inside Copilot in Word, Excel and PowerPoint for eligible Frontier preview customers. A dedicated administrator setting is disabled by default. EU, EFTA and UK customers are excluded during preview; SpaceXAI is a listed subprocessor.
Limits and interpretation: Microsoft controls admission, defaults, eligibility, orchestration and distribution. Model choice is not an open marketplace or a right to leave Copilot. Preview is not general availability.
NSF awarded $35 million over five years to the San Diego Supercomputer Center, working with the Texas Advanced Computing Center, to establish the National Artificial Intelligence Research Resource Operations Center. The center will operate the national portal, integrate computing, data, models and tools from public and private providers, and provide support, training and coordination as NAIRR transitions from a pilot toward sustained infrastructure.
Limits and interpretation: A substantial portion of frontier capability beneath that public interface remains supplied by a concentrated commercial industry. NAIRR therefore increases contestability without eliminating dependence on concentrated upstream providers or proving full workload portability among them.
GSA and OpenAI executed a 27-month OneGov agreement, effective October 1, 2026 through December 31, 2028, that makes ChatGPT available to eligible federal, state, local and tribal governments with a 50% discount on consumption-based model use, no platform-access fee, no minimum order or spending commitment, multiple purchasing routes, government-data safeguards, training and deployment support. The agreement also includes most-favored-pricing protections, utilization and savings reporting, and a one-year checkpoint.
Limits and interpretation: The agreement was executed in September but takes effect October 1. September status is adopted procurement access, not completed implementation.
GSA’s USAi platform completed its transition from a free pilot into a paid cost-recovery shared service for federal agencies. USAi is already used by 25 agencies, with 16 more expected to onboard by year-end, and provides a shared Chat interface, unified API and administrative/evaluation Console across models from multiple providers. Agency charges combine a platform fee with model usage passed through at cost.
Limits and interpretation: The service remains federal-only, agencies must now fund their own participation, and substantial authority over provider selection and platform architecture is concentrated in GSA. The new cost-recovery model could also produce uneven adoption across agencies with different budgets and implementation capacity.
Action/date record: September 2026; processed September 14. Approved score: Dispersion +1. Primary lane: 1. Secondary lanes: 5.
Microsoft opened a formal process through which outside participants can influence the behavioral rules for MAI. Under the project’s consultation-as-institution rule, the operative participatory channel is scoreable now even though the resulting code and its enforceability remain unsettled.
Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.
September scored actionSeptember 30, 2026 evidence cutoff
Action/date record: September 2026; processed September 15. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: 7.
Salesforce and Google implemented Model Context Protocol connections across their agent and enterprise environments. The score recognizes a functioning standardized interface that lowers one layer of integration friction between major ecosystems; it does not establish full workflow, identity, memory or data portability.
Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.
September scored actionSeptember 30, 2026 evidence cutoff
Action/date record: September 2026; processed September 16. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: 4.
Google exposed Home capabilities through MCP, allowing outside agents to interact through a standardized tool layer rather than only through Google’s own assistant. The score remains +1 because Google retains platform authorization, device access and execution control.
Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.
September scored actionSeptember 30, 2026 evidence cutoff
Amazon prevented Meta's Muse agent from using Amazon's shopping infrastructure, preserving Amazon's control over an important commerce surface and limiting an outside assistant's ability to complete transactions there.
Source hold: The approved ledger refers to an alert source archive, but an operative URL was not recovered from the records materialized for this review. The score is preserved; actor, precise action date, text and scope need source-archive reconciliation before this becomes fully verified public evidence.
September scored actionSeptember 30, 2026 evidence cutoff
Xiaomi released full MiMo-V2.6 Pro-RL and Flash-RL checkpoints, together with a 9B distilled model, under the MIT License. The family supports multimodal inputs, tool use, a one-million-token context and independent deployment through common open runtimes.
Action/date record: September 22, 2026 announcement; processed September 23. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: 7.
Public preview provides managed persistent workspaces and governed tool access for competing agent harnesses and custom agent software. DigitalOcean identifies customers using the service.
Limits and interpretation: Preview is not general availability. Managed state, permissions and gateway services can create new dependencies; easy migration of accumulated work is not established. Performance and productivity figures are vendor-reported.
Action/date record: September 30, 2026. Approved score: Dispersion +1. Primary lane: 1. Secondary lanes: 5. Stable record identity: D1001-24.
SB1000 chapter 861 and AB2713 chapter 856 were approved and chaptered September 30. SB1000 removes the one-million-monthly-user provider threshold, requires publicly accessible provenance-verification tools and latent metadata for content creation or material alteration, and changes existing transparency duties. It also removes the optional visible-label provision and delays specified assistive-technology application. AB2713 revises platform preservation, detection and display of system provenance, with privacy, copyright, technical-feasibility and qualifying interoperable/widely adopted standards conditions. The platform provisions operate January 1, 2027; SB1000 is an urgency statute effective on enactment.
Limits and interpretation: Combined scope review supports a modest +1 because broader provider coverage and specified provenance access/preservation obligations increase institutional visibility. Some provisions narrow or qualify duties, and platform mechanisms build on existing law. Do not describe all amendments as strengthening or count the underlying act again.
Action/date record: September 29, 2026 filing; publicly released September 30. Approved score: Concentration +1. Primary lane: 1. Secondary lanes: 5. Stable record identity: D1001-01.
The Third Circuit’s substantive copyright/fair-use opinion sustains protection for independently edited Westlaw headnotes and rejects fair use for training a competing non-generative legal-search product. Underlying judicial opinions remain freely available. The decision is a merits ruling on partial summary judgment reached through an interlocutory appeal, rather than temporary relief while merits remain undecided.
Limits and interpretation: Control over edited legal-information inputs constrains a competitor’s training pathway. A narrow +1 recognizes this institutional consequence without claiming a ban on all AI training or all generative systems. Earlier scored Thomson Reuters proprietary-model launch is a distinct event.
Action/date record: August 25, 2026; recorded in the September log when the structural significance was identified. Approved score: Dispersion +1. Primary lane: 5. Secondary lanes: none recorded.
WhatsApp established a separate Third Party Agent Platform through which users can connect AI agents built, owned and operated by independent providers rather than Meta. The current terms allow users to connect up to five third-party agents, and WhatsApp's Help Center states that users can chat with third-party agents, including agents they create through outside tools and code.
Limits and interpretation: The rollout is still limited, Meta retains platform control and can change limits or suspend access, and the older Business Solution restriction on general-purpose AI providers remains in place. The new framework therefore creates meaningful separability without making WhatsApp a fully neutral AI platform.