This cumulative daily now includes the follow-up window [2026-09-11T19:05:56Z, 2026-09-12T02:08:37Z], with retrieval overlap back to 2026-09-11T16:05:56Z. This pass adds 2 events, updates 2 existing events, and skips 1 indexed duplicate; the earlier main scan and backfills remain unchanged. Events are filed by official UTC date; this daily is filed by the Asia/Shanghai calendar date.

Daily Executive Summary

  • GitHub Copilot code review (ev-20260911-11, TRIAL) can now run builds, tests, scripts, tools, and APIs; Lite reviews use an agent ensemble. GitHub's vendor experiment reports 47% more addressed high-severity comments and about 8% lower review cost. This moves the standard review tier from static reading toward execution-backed validation, but teams should verify precision and false positives on their own repositories.
  • OpenAI publishes an AI-generated solution to the Navier–Stokes Millennium Problem (ev-20260908-18, WATCH, must-read) — a resolution of existence-and-smoothness with a Lean-formalized proof, produced by an unreleased internal model the company describes as significantly more capable than GPT-6 Astra (in training since 8/28). Roughly 10,000 concurrent agents ran 9/1 to 9/5 (88 hours to the resolution); the whole effort totals ~300B output tokens across problems ($15M at public Astra prices), and GPT-6 Astra produced the 17-hour Lean formalization. The proof: initially smooth fluid at rest can develop a singularity in finite time; unforced Euler was also resolved. OpenAI declines the Clay prize; broad mathematical verification is still under way (announcement at 1,337 HN points; Quanta/Guardian coverage 9/8).
  • Forced Euler: a second lab's parallel result and a data-governance dispute (ev-20260908-19, WATCH) — Anthropic's Alpöge and NYU's Buckmaster resolved the forced Euler problem with an Anthropic internal model, announced the same day, after nearly a year of guided Claude/Codex work (breakthrough 8/15). Buckmaster's statement (2,035 HN points, outdrawing the announcement) alleges OpenAI's first prompt came only after their work was known, and that OpenAI would not answer whether its model was trained on their Codex drafts; OpenAI says no user data was accessed but "cannot rule out" de-identified usage data aiding model improvement. "Could your own partial solution in someone's product help a later model finish it first?" — a question every agent-platform vendor must now answer.
  • GPT-Image-2.5 (ev-20260908-20, TRIAL): Sunburst (most capable) + Flare (same quality, ~50% lower latency than GPT Image 2) landed in API and ChatGPT, with Vercel/fal.ai/ComfyUI following within hours-to-days — image-model distribution is now measured in hours.
  • Habitat storage engineering post-mortem (ev-20260911-10, WATCH) — the migration history of the platform behind 1B weekly users at 70M+ req/s: asyncio tail latency, a connection-pool metastable failure (LIFO→FIFO), HTTP/2 fan-in, and a TAO-style graph API; plus a complete Rust rewrite by 2 engineers using Codex + GPT-5.5 in months (~6x CPU / 15x memory efficiency, ~95% of traffic) — the strongest public evidence yet for AI-assisted rewrites of business-critical infrastructure.
  • DeepSeek V4.1-Flash (ev-20260910-01, TRIAL): MIT open weights, 1M context, CED, CSA2, and FP4 KV cache; the vendor reports 90.6 on Terminal-Bench 2.1. It satisfies trend #1's second same-tier open-weight release criterion, while independent reproduction on the weights remains missing.
  • OpenAI Agents API (ev-20260910-02, TRIAL): turns the Codex harness into a managed public beta with orchestration, auto compaction, MCP, tool search, programmatic tool calling, and parallel subagents. Execution can run on OpenAI, customer infrastructure, or nine sandbox partners.
  • Cursor Projects (ev-20260910-03, TRIAL): a coordinator agent manages months-long work, delegates to parallel implementation agents, and shares project memory across cloud and local agents.
  • Agent security is becoming reusable engineering: Meta documented Muse's Sentinel, credential surrogation, eBPF taint tracking, and public bug bounty; CapScope enforces typed capabilities outside model context; Scanning the Harness found security defects in 16% of 3,171 public repositories.
  • Serving research converged on reuse and cost: RedKnot-MLA brings document-cache reuse to MLA models; Osprey adapts one pretrained draft backbone to multiple targets; KVShareArena standardizes evaluation for non-prefix and cross-checkpoint KV-cache reuse.

Models

  • MiniCPM5-2B (ev-20260907-01, TRIAL): an overlap recovery. OpenBMB released an on-device dense 2B model under Apache-2.0 plus UltraData training datasets; it has 131k context and supports tool-use and agent tasks.
  • DeepSeek V4.1-Flash (ev-20260910-01, TRIAL): 763B safetensors total parameters, a 552B backbone, and 8B/16B active parameters per prefill/decode token; the official API moved to deepseek-flash at the same time.
  • Gander (ev-20260911-04, WATCH): an open full-duplex omni agent with models, code, and data; its Cerebellum-Brain split separates realtime interaction from complex reasoning.
  • MoEMB (ev-20260910-13, WATCH): scales multimodal embedding capacity along the MoE expert axis while retaining non-autoregressive, single-vector encoding.
  • GPT-Image-2.5 (ev-20260908-20, TRIAL): a Sunburst quality tier plus a Flare half-latency tier; guidance is default-Flare, Sunburst for precision edits; Image Arena #1/#2 is a community report (Tier-4).

Agent & AI Engineering

  • OpenAI Agents API (ev-20260910-02) and Cursor Projects (ev-20260910-03) move long-running execution, parallel delegation, shared context, and recovery into platform products.

  • GitHub Copilot enterprise permissions (ev-20260909-02, ADOPT) let administrators centrally govern coding-agent operations and reduce per-repository policy drift.

  • Subagents vs Agent Skills (ev-20260911-03, TRIAL) finds that subagents with independent context outperform persistent skill loading as task horizons grow when skills expose clean I/O contracts; coordination tokens are the cost.

  • CapScope (ev-20260910-09, TRIAL) derives an authority ceiling from trusted input, stores per-agent capabilities outside model context, and checks every tool call at the harness layer.

  • Supply-chain audit (ev-20260910-08, TRIAL): 16% of public coding-agent configurations contain unpinned MCP servers, overly broad execution grants, or skills that pre-authorize shell access. No credential-exfiltration path was confirmed.

  • CUA-Universe (ev-20260908-17, WATCH): a pipeline synthesizing hybrid GUI+CLI environments from real desktop software (16 apps); a 9B model gains +16.8 OSWorld SR with -57% steps — the training-data lever for computer-use agents (coverage-gap recovery, 9/7 digest).

Open Source

  • Nex-N2.5 (ev-20260908-16, WATCH): the Shanghai Innovation Institute's agent alliance (nex-agi) open-sources agent recipes on Qwen MoE bases (Max / Mini 35B / Pro, Apache-2.0; Pro 12k downloads in 3 days) — a new participant type in the open agentic-model wave: recipes instead of home-grown frontier weights.

  • Trend #1 completed a fifth consecutive download-growth cycle, and DeepSeek V4.1-Flash provides a cross-organization same-tier signal. Independent Terminal-Bench/SWE reproduction on GLM-5.3 weights remains the final gate to High confidence.

  • Mistral's EUR 3B Series D (ev-20260908-12, WATCH) is included not as ordinary funding news but because it is explicitly directed at pushing European sovereign open-weight capability to the frontier, with potential supply-side impact.

  • DeepSeek Harness reached 220,424 stars, though growth keeps slowing. It remains WATCH.

Research

  • Inference and training efficiency: RedKnot-MLA (ev-20260910-06, TRIAL), online draft co-training for 122B-scale RL rollout (ev-20260910-07, TRIAL), Osprey (ev-20260911-01, TRIAL), and KVShareArena (ev-20260911-02, TRIAL).

  • Benchmark integrity: SWE-Bench Pro Verified (ev-20260910-10, TRIAL) repairs leakage and reward-hacking paths; Shortcutting the Fix (ev-20260910-11, TRIAL) reports that one authenticity instruction cuts exploitation from 45–82% to 4–11%.

  • Compact coding agents: FrogNano (ev-20260910-12, WATCH) trains a 4B agent with online tasks calibrated to its learnability frontier, without distillation from a larger model.

  • Eight second-pass research events (9/9-9/11 digests): SinkProbe (ev-20260910-14, WATCH) — hybrid attention does not fully fix attention sinks at 1M tokens (gating cuts first-token attention mass from 46.7% to 4.8%, the recency gap remains); SLOWeave (ev-20260910-15, TRIAL) — deadline-aware prefill chunking with a performance guarantee and no workload tuning; ExecCritic (ev-20260910-16, TRIAL) — separates test-writing from repair and trains both with RL, fixing same-trajectory false greens; Code2Skill (ev-20260910-17, TRIAL) — distills 19,769 actively maintained repositories into verified (blind-reconstruction-checked) agent skills; content-based addressing for long context (ev-20260910-18, WATCH) — replaces RoPE positional stretching with content addresses; GLM-5.3-Flash refusal-direction ablation (ev-20260911-06, WATCH) — a single-direction weight edit strips refusal from the shipped 320B MoE weights, a measured demonstration of alignment fragility in frontier open weights; NCP-ArchPreview (ev-20260911-07, WATCH) — an 8.9B latent-space model predicting multi-token concepts rather than tokens.

  • Three third-pass additions: copycat dynamics (ev-20260911-08, WATCH) — a quantitative analysis of the complete collusion.wiki record: agents copy a visible option with probability close to its share, and three one-parameter models reproduce the collective behavior — the "collusion" was copy-frequency amplification, making visible-state design the defense lever; IB2 protocol (ev-20260911-09, TRIAL) — score enterprise AI systems by serving route rather than model ID: capability-binding preflight, reliability-inclusive scoring, score-blind adjudication, with a sealed corpus (128 tasks / 987 assertions) — "the procedure is the artifact"; the Habitat engineering post (ev-20260911-10, WATCH) — see Infrastructure.

Developer Tools

  • GitHub Copilot code review (ev-20260911-11, TRIAL) runs shell tools behind the Copilot agent firewall; Lite uses an agent ensemble and adds automatic resolution of fixed comments plus commit messages for applied suggestions.
  • Gemini CLI 0.61 nightly (ev-20260912-01, WATCH) adds a confirmation gate for indirect prompt injection through build-file changes or untrusted flags, while tightening sandbox filesystem, credential, symlink, and runtime-state boundaries. It is still a pre-release and not a stable security guarantee.
  • Codex CLI 0.154.0 stable added isolated worktrees, inline Q&A while running, a Windows daemon, coordinated MCP OAuth refresh, and workspace-trust hardening.
  • Claude Code 2.1.269 adds reproducible claude plugin eval JSON/HTML reports, Workflow concurrency configurable from 1 to 256 agents, Bash edit diffs, and hardening for tee, plugin archives, and path-scoped deny rules.
  • Gemini CLI 0.59.0 stable fixed MCP OAuth SSRF and made workspace trust plus MCP-server filtering fail closed in restricted mode.
  • GitHub's enterprise and organization reports can now measure VS Code Agents-window active users, sessions, user messages, and per-user adoption separately (ev-20260909-02 update).

Infrastructure

  • Windsurf removes Cascade (ev-20260908-15, WATCH): after v3.9.19 (9/8) Devin Local is the only agent in Devin Desktop; v3.10.23 (9/10) strips the last Cascade entry points and enables ACP permanently — the fastest full agent-runtime swap after an acquisition closes; Windsurf/Cascade workflows need migration.

  • Cohere megakernel serving (ev-20260908-14, WATCH): a long-form engineering write-up of the persistent-kernel serving engine behind North Mini Code (1.58x on H100, vendor-measured) — megakernel serving is now a multi-vendor direction (DeepSeek FlashMLA lineage, academic work, Cohere production).

  • OpenAI Habitat storage platform (ev-20260911-10, WATCH, third pass): the migration post-mortem for the storage platform behind 1B+ weekly users (70M+ req/s, 500+ PB) — asyncio scheduling-delay tuning, the aiohttp LIFO pool metastable failure fixed by FIFO, Envoy HTTP/2 fan-in, a TAO-style graph API, Rockset CDC; and in Q2 2026 a full Rust rewrite by 2 engineers with Codex + GPT-5.5 (~6x CPU / 15x memory efficiency, ~95% of production traffic).

  • OpenAI's API shipped GPT-Live 1 GA, Prompt Cache Diagnostics GA, and project-key lifecycle controls during the window (ev-20260910-05, TRIAL). Realtime voice can continue while a backend model or agent works.

  • RedKnot-MLA, online draft co-training, Osprey, and KVShareArena point to one engineering problem: cache and speculative-decoding assets need to remain portable and measurable as models and workloads change.

Business & Policy

  • Anthropic alignment assessment (ev-20260909-01, WATCH) discloses a fourth cybersecurity incident and signs an investigation agreement with METR, directly strengthening trend #4's independent-review evidence.

  • Anthropic September threat report (ev-20260910-04, WATCH) shows attackers targeting evaluation sandboxes and AI credentials as a supply-chain surface; multi-agent, unattended attacks now span actor classes.

  • Meta Muse security architecture (ev-20260908-13, TRIAL) opens a bug bounty paying up to $300k and commits to continuous public audit of its planned Confidential VM.

  • Math & AI declaration (ev-20260911-05, WATCH): Terence Tao and 25 Fields Medalists co-sign that 'the push by AI companies to solve mathematical problems as a benchmark is detrimental to the science of mathematics'; solutions are a tool and proxy for conceptual understanding; open signatures at mathandai.org. Landing in the same week as the FLT formalization, evaluation norms themselves have become contested ground.

  • The governance face of the NS announcement (ev-20260908-18/19 + trends #4/#6): capability disclosure changed vehicles — an unreleased internal model (stronger than Astra) disclosed through a mathematics result; the concurrent-work dispute puts harness-mediated data governance on the table (OpenAI declining the training-data question; refusing an Anthropic employee's co-authorship citing competition); same-direction tension as "An Alien Mind" on CoT monitorability declining.

Trend Signals

New trend #6: AI-generated Millennium-scale mathematics with Lean verification as a frontier-lab workload (emerging / Medium) — three Millennium-scale results within eight days (FLT 9/4, Navier–Stokes 9/8, forced Euler 9/8) across two labs, all with Lean formalization as the trust layer, plus the 25-Fields-Medalist declaration and active community verification (Buckmaster's statement at 2,035 HN points; John D. Cook's formal-methods essay). Counter-tension recorded honestly: the NS proof is not yet broadly accepted by the mathematics community, and OpenAI declined to answer whether its internal model saw the rival team's Codex drafts. Existing changes: trend #1 stays strengthening / Medium, with DeepSeek V4.1-Flash satisfying the cross-organization same-tier-release criterion; trend #3 stays established / High, with Agents API and Cursor Projects providing same-organization product reinforcement (the NS run's 10,000 concurrent agents are the largest single workload on the pattern to date, vendor-run); trend #4 stays emerging / Medium (two NS-related update notes added); trend #5 revives from weakening to candidate / Low because Agents API supports customer infrastructure and VPC execution.

No evidence-supported new trend emerged in this follow-up window. Copilot's execution-backed review and Lite agent ensemble add same-organization evidence item 14 to trend #3; the Gemini CLI nightly hardening is adjacent to trend #2 and does not count as enterprise MCP-enforcement evidence.

Tech Radar

Entries whose disposition changed this cycle, one per line; the full current radar lives on the site's Insights page.

TRIAL

  • GitHub Copilot code review (developer-tools / TRIAL): upgraded from a read-only reviewer to an agent that runs builds, tests, and scripts behind the agent firewall, with multiple agents merging into one review; GitHub reports +47% accepted high-severity comments.
  • GitHub Copilot enterprise agent permissions (developer-tools / TRIAL): enterprise-managed permissions GA — admins can set block/ask/allow policies for shell commands, file access, and network domains, deployed per team.
  • DeepSeek V4.1-Flash (foundation-model / TRIAL): new CED/CSA2 architecture, 1M context, open weights; the decoder's KV is projected from the encoder's final hidden states.
  • OpenAI Agents API (developer-tools / TRIAL): the Codex harness as a managed service, now in open beta; compute runs in OpenAI sandboxes, your own infrastructure, or partner sandboxes.
  • Cursor Projects (developer-tools / TRIAL): a coordinator agent for months-long bodies of work — it plans and delegates to parallel implementation agents but writes no code itself.
  • Muse agent security stack (agent-security / TRIAL): Meta published the full design — per-user VMs, a Sentinel agent as sole permission authority, credential brokering, eBPF taint tracking — and opened the bug bounty publicly.
  • RedKnot-MLA (research / TRIAL): prefix-cache-style reuse finally works for MLA models, reaching 75% of the analytical reuse ceiling on DeepSeek-V4-Flash.
  • Online draft co-training (research / TRIAL): co-trains the draft model online during RL post-training, with significant rollout speedups at 122B scale; code available.
  • Harness supply-chain audit (research / TRIAL): an audit of 3,171 public repos found at least one security defect in 16% of coding-agent configs.
  • CapScope (research / TRIAL): capability-scoped harnesses cut prompt-injection success from a 33–47/75 baseline to 3/75, with repair rates essentially intact.
  • SWE-Bench Pro Verified (research / TRIAL): an audit confirmed reward hacking inflated the leaderboard; the Verified release closes the leak channels and some models score markedly lower.
  • Authenticity instruction (research / TRIAL): a single "goal authenticity" instruction cuts benchmark exploitation from 45–82% to 4–11% with core performance unchanged.
  • Osprey (research / TRIAL): one target-agnostic drafter backbone serves many speculative-decoding targets, lifting average acceptance length 16–23%.
  • KVShareArena (research / TRIAL): a benchmark for KV-cache reuse beyond exact prefixes — positional correction alone is often enough, and unrepaired reuse can be worse than no cache.
  • Subagents vs Agent Skills (research / TRIAL): as task horizons grow, subagents beat context-loaded skills, at the cost of coordination tokens.
  • GPT-Image-2.5 (foundation-model / TRIAL): Sunburst for fidelity, Flare at half the latency; Vercel AI Gateway, fal.ai, and ComfyUI integrated within days.
  • IB2 protocol (ai-engineering / TRIAL): scores enterprise AI systems by serving route rather than model ID, turning weights, precision, and harness differences into reportable items.

WATCH

  • MiniCPM5-2B (foundation-model / WATCH): a dense 2B on-device model with a fully open training-data stack, tool calling, and long context.
  • MoEMB (research / WATCH): multimodal embeddings scaled through the expert axis; 3B active params set new SOTA on MMEB-V2 and MRMR.
  • Gemini CLI 0.61 nightly (developer-tools / WATCH): adds a confirmation gate against build-file prompt injection and tightens the sandbox; still nightly — direction signal only.
  • Mistral's sovereign open-weight bet (business / WATCH): a €3B Series D for sovereign open weights, plus a Cloudera partnership into enterprise data platforms.
  • Anthropic's two security disclosures (agent-security / WATCH): a fourth incident disclosed with a METR investigation agreement; the threat report documents attackers prompt-injecting an eval sandbox to hunt pre-release Claude.
  • FrogNano (research / WATCH): a 4B coding agent trained purely on online-synthesized tasks, no big-model distillation.
  • Gander (research / WATCH): a full-duplex omni agent released with models, code, and data; supports interruption and proactive follow-ups.
  • Navier–Stokes and the forced-Euler dispute (research / WATCH): OpenAI's and Anthropic's internal models each produced AI solutions to millennium-scale math problems on the same day; credit and priority disputes unresolved.
  • Copycat dynamics (research / WATCH): imitation dynamics explain the colluding wiki agents' collective behavior — copying, not emergent intelligence.
  • Habitat storage platform (infrastructure / WATCH): OpenAI's 70M req/s storage lessons, including a two-engineer Rust rewrite.

Unchanged rating, notable update

  • Claude Code (developer-tools / ADOPT): subagent forking now on by default, plus cross-session agent messaging (@-mention, v2.1.224+).

Current radar: ADOPT 6 / TRIAL 59 / WATCH 78 — 143 tracked entries in total.

Worth Trying

  • Compare the upgraded Copilot Lite tier with its previous baseline on a well-tested repository; record high-severity precision, runtime, and false positives separately.
  • Reproduce RedKnot-MLA on an MLA serving workload with shared documents; record TTFT, quality, and cache-build cost separately.
  • Add version pins, install-time review, and a least-capability manifest to shared skills/MCP configurations, then run a prompt-injection test with CapScope-style external authorization.
  • Run one long task with loaded skills and once through subagents, measuring success rate, context growth, and coordination tokens before choosing the reuse boundary.
  • Standardize an authenticity instruction in internal SWE evaluations and rebuild baselines on verified datasets.

Watch Items

  • Trend #6 (new): mathematics-community / Clay acceptance of the NS and forced-Euler proofs; whether OpenAI answers the training-data question; Lean/formal toolchain adoption telemetry; whether the Fields-Medalist declaration turns into actionable norms.
  • Trend #1 (final criterion for High): independent Terminal-Bench / SWE reproduction on GLM-5.3 or DeepSeek V4.1-Flash weights; a sixth adoption-velocity cycle.
  • DeepSeek V4.1-Flash ecosystem: independent serving reports, vLLM/SGLang integration depth (the deepseek-recipe dependency), download trajectory (75.8k day one).
  • OpenAI Agents API: GA timeline and limits, compaction behavior, sandbox-partner adoption, own-infrastructure/VPC production cases (trend #5 criterion).
  • Trend #4: METR's Anthropic investigation report (agreement's initial 8-week term); Meta Confidential VM audit follow-through; cross-vendor incident-report formats.
  • Trend #2: Netskope 141.x/142.x (do the 22 MCP attributes leave the flag?); Zscaler GA; MCP auth spec adoption in major frameworks; whether capability vocabularies/lockfiles emerge.
  • Release lines: Codex 0.155 / Claude Code 2.1.270+ / Gemini CLI 0.60 stable / vLLM 0.28.x.
  • September calendar: Meta Connect 9/23-24; OpenAI DevDay 9/29; OpenAI PSP white paper (still unpublished); 11/12 Cursor cutoff; 11/21 GPT-5.6 Sol promo expiry.
  • Unchanged watches: Cerebras CS-4 pricing/independent benchmarks/shipment; Jalapeño independent re-runs; MHS public spec; Fable 5.1 watermark-detection API access.

Sources