AI Intelligence Trends - 2026-09-23

No new trend has enough independent evidence in this cycle.

Trend #1: Chinese labs' open-weight agentic coding models compete at the frontier

  • Status: strengthening / Medium
  • First observed: 2026-08-15
  • Last updated: 2026-09-23
  • Evidence: vLLM 0.30 moves several serving paths for DeepSeek V4.1, GLM-5.3 and other open-weight models into a tagged stable runtime. Third-party Terminal-Bench or SWE reproduction is still missing.
  • Why it matters: Self-hosted frontier coding still needs independent evaluation on public weights.
  • What would confirm: Third-party Terminal-Bench or SWE runs on the weights.
  • 2026-09-23 review: serving-adoption evidence strengthens, but the final public-weight reproduction gate for High confidence is still unmet. Stays strengthening / Medium.

Trend #2: MCP enters enterprise security and enforcement

  • Status: emerging / Medium
  • First observed: 2026-08-15
  • Last updated: 2026-09-21
  • Evidence: No new cross-vendor auth/enforcement specification, GA security product or public adoption data appeared.
  • Why it matters: MCP governance spans identity, policy, telemetry and lifecycle.
  • What would confirm: A second security vendor's GA detection product and an interoperable auth spec.
  • 2026-09-23 review: stays emerging / Medium.

Trend #3: Coding agents converge into multi-agent runtimes

  • Status: established / High
  • First observed: 2026-08-15
  • Last updated: 2026-09-23
  • Evidence: RRSI adds proposal budgets, held-out tasks and pruning to automated agent-harness optimization; a long-horizon collusion study shows reward conflicts can reshape multi-agent cooperation. Both are runtime research signals, not new production adoption.
  • Why it matters: The selection unit is now model, runtime, scaffold, monitoring and delegation topology.
  • What would confirm: More independent production cases and convergence in orchestration interfaces.
  • 2026-09-23 review: evidence on harness management and long-running coordination risk increases. Stays established / High.

Trend #4: Frontier labs institutionalize incident disclosure and independent review

  • Status: strengthening / Medium
  • First observed: 2026-08-26
  • Last updated: 2026-09-19
  • Evidence: No METR findings, evaluator-independence rules or cross-vendor disclosure format appeared.
  • Why it matters: Trust is moving from vendor claims to standing access, investigable evidence and third-party review.
  • What would confirm: METR findings, explicit evaluator independence and a cross-vendor disclosure format.
  • 2026-09-23 review: stays strengthening / Medium.

Trend #5: Enterprise self-hosted and data-resident execution planes form

  • Status: candidate / Low
  • First observed: 2026-08-18
  • Last updated: 2026-09-15
  • Evidence: vLLM 0.30 expands self-hosted inference, but no new enterprise agent execution plane, independent production adoption, OpenAI PSP paper or Anthropic EFS delivery appeared.
  • Why it matters: Execution location is becoming an enterprise procurement axis.
  • What would confirm: Independent production adoption, OpenAI PSP and Anthropic EFS delivery.
  • 2026-09-23 review: stays candidate / Low.

Trend #6: AI-generated millennium-scale mathematics and Lean verification become a frontier workload

  • Status: emerging / Medium
  • First observed: 2026-09-04
  • Last updated: 2026-09-15
  • Evidence: No new millennium-scale mathematics result, Lean formalization or independent acceptance appeared.
  • Why it matters: Formal verification gives long-running agent output a stronger trust layer.
  • What would confirm: Formal acceptance, third-party reproduction and Lean adoption data.
  • 2026-09-23 review: stays emerging / Medium.

Trend #7: Agent runtimes separate broad capability from consequence control

  • Status: emerging / Medium
  • First observed: 2026-09-09
  • Last updated: 2026-09-23
  • Evidence: Across 14 models and eight domains, DUMA-Bench finds that letting users keep changing shared state raises attack success from 26.9% to 41.1%. Risk depends on the model, user and state changes permitted by the runtime.
  • Why it matters: Generated commands and permitted consequences should be independently testable and auditable.
  • What would confirm: One policy schema across shell, MCP and browser with public adoption telemetry.
  • 2026-09-23 review: new evidence strengthens the need for two-layer acceptance tests; a unified policy schema is still missing, so the trend stays emerging / Medium.