AI Trend Snapshot - 2026-09-26

Based on the scan window from 2026-09-24T16:01:46Z to 2026-09-25T16:01:00Z. Numbering follows the stable order in trends/current.json.

No evidence-supported new trend emerged in this cycle. New signals reinforce existing directions in agent runtimes, continuous evaluation and consequence control.

Trend #1: Open-weight agentic coding models from Chinese labs competing at frontier level

  • Status: strengthening / Medium (maintained)
  • Review: No third-party Terminal-Bench or SWE reproduction on public weights appeared. LFM2.5-VL-DSpark is a speculative-decoding component and does not meet this trend's coding-capability criterion.
  • What would confirm: A third party reproduces Terminal-Bench or SWE-class results on public weights.

Trend #2: MCP entering enterprise security & enforcement phase

  • Status: emerging / Medium (maintained)
  • New evidence: Google Cloud API Gateway can expose OpenAPI REST operations as MCP tools while reusing JWT, quotas and logs; GitHub's global default brings MCP-server policy into enterprise governance.
  • Counter-signal: tools/list is unauthenticated by default and API keys cannot protect discovery, showing that cross-platform secure defaults remain immature.
  • What would confirm: A shared authorization and audit schema across vendors plus independent production adoption.

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

  • Status: established / High (maintained)
  • New evidence: Microsoft Foundry connects durable execution, A2A, Routines, tool search and production optimization; Android Studio hosts replaceable agents through ACP; Codex 0.157, Qwen Code 0.24.5 and Claude Code 2.1.282 strengthen background sessions, traces and policy reliability.
  • Adjacent evidence: Nubank's production simulation and HEXIS state machines show runtime differentiation moving from tool access to qualification and control of long-running behavior.
  • What would confirm: Interoperable task, identity, approval and trace schemas across products.

Trend #4: Frontier labs institutionalizing safety-incident disclosure and independent third-party review

  • Status: strengthening / Medium (maintained)
  • New evidence: Nubank validates simulation against live A/B tests, Microsoft converts production traces into evaluation and optimization data, and Modelun publishes low-cost cross-release assays with judge-agreement materials.
  • Assessment: These signals strengthen continuous evaluation delivery, but they are not a new independent-review institution at a frontier lab.
  • What would confirm: A second frontier lab adopts routine independent review, or vendors converge on one disclosure schema.

Trend #5: Enterprise self-hosted / data-residency execution planes forming across AI toolchains

  • Status: candidate / Low (maintained)
  • Review: No new self-hosted execution plane, independent production adoption, OpenAI PSP whitepaper or Anthropic EFS deployment appeared. Microsoft Foundry is a managed-platform expansion, not self-hosting evidence.
  • What would confirm: At least two independent production deployments and delivery of PSP/EFS commitments.

Trend #6: AI-generated Millennium-scale mathematics with Lean verification as a frontier-lab workload

  • Status: emerging / Medium (maintained)
  • Review: No new Millennium-scale result, Lean formal proof or independent acceptance appeared. ExplorationBench is a verifiable-exploration benchmark, not mathematics-proof evidence.
  • What would confirm: Formal mathematical acceptance, independent reproduction or Lean-toolchain adoption data.

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

  • Status: strengthening / Medium (maintained)
  • New evidence: Trace tampering and EvasionBench expose audit-record and model-monitor boundaries; Google API Gateway MCP, GitHub default policy and Qwen attestation move control further into platforms and runtimes.
  • Assessment: Independent append-only trace capture and cross-turn consequence gates should become acceptance criteria. The two security papers share one research team, so they do not independently establish a new multi-actor trend.
  • What would confirm: An interoperable policy schema spanning shell, MCP and browser execution plus public adoption telemetry.

Invalidated / retired

(None)