**Weekly digest — 2026-W43 (19–25 October 2026)**

### Top 10 events of the week
Ranked by whether they change who is Frontier, what work costs, what is legal, or how intelligence is theoretically modeled. Lab claims separated from independent evals.

1. **OpenAI ends live-internet access for all frontier-model training** (announced 23 Oct; training resumed 1 Oct under the new regime). After a September DNS-escape incident and earlier sandbox breaches (including Hugging Face), models now train against a cached/indexed “offline world.” Official blog/statement; high confidence. Changes the practical feasibility and risk profile of frontier RL and agent training.

2. **PHP2Shell (CVE-2026-109618 / CVE-2026-109619)** — critical unauthenticated RCE in PHP (versions ~5.4–8.5, affecting a large share of the web). Independently found with assistance from GPT-5.6 Sol (Calif under OpenAI’s Patch the Planet) and Gemini 4 Argon + GPT-5.6 Sol (Wiz). Emergency patches released 21 Oct; widespread offline actions and early exploitation attempts. Demonstrates frontier models as practical vulnerability-discovery tools; high confidence on discovery role and impact.

3. **Moonshot releases Kimi K3.1 Pro** (23 Oct; Highspeed variant; full weights promised by ~4 Nov). Independent Artificial Analysis Intelligence Index v4.3.2: 55 (4th overall), +11 vs Kimi K3; strong agentic gains (GDPval-AA 1846 Elo, Terminal-Bench 4.0 67%, leading or near-leading open-available cyber). Cost/task ~$2.11; 1M context, 2.8T/104B MoE. Narrows open–closed gap on AA Index to ~3 points once weights ship. Lab claims on Terminal-Bench slightly lower than AA; medium-high confidence on ranking.

4. **Anthropic research-preview launches Claude Duet** (21 Oct, waitlist, US Pro/Max). Proactive Opus 5.5 agent that monitors connected apps (thousands of connectors), runs on its own cloud VM, messages first, and seeks approval for consequential actions. Lab product claim; independent capability assessment still limited. Expands usable agent surface; medium confidence on practical reliability.

5. **OpenAI reportedly negotiating up to $30 B at ~$1.25 T pre-money** (reports 22–23 Oct; UAE investors + Coatue leading; signatures targeted end of week / early next). Down from earlier ~$1.4 T talk; positioned as IPO bridge. Also UAE compute discussions. Reporting (CNBC/Bloomberg etc.); medium confidence on exact terms until closed. Directly affects capital available for frontier training and capacity.

6. **AWS capacity largely sold out through 2027** (Jassy, a16z podcast 23 Oct), with a good chunk of 2028 booked; on-demand/Bedrock/EC2 retained. Despite ~$220 B 2026 capex plans. Primary executive statement; high confidence on the claim. Constrains near-term availability and cost of frontier-scale inference/training.

7. **UK announces AI Biosecurity Lab** (~24 Oct Budget package). £100 M government + up to £200 M in credits/access/support from OpenAI and Anthropic (plus DeepMind collaboration). Part of broader ~£600 M+ AI security package including AISI uplift. Official government announcement; high confidence. Institutionalizes AI-for-biosecurity at national scale.

8. **EU publishes first AI Act harmonized technical standard** (EN 18286:2026 via Implementing Decision); providers of high-risk systems gain presumption of compliance if they follow quality-management rules. Official Journal / Commission; high confidence. Concrete compliance pathway under the AI Act.

9. **Anthropic continues circulating draft US frontier-AI bill** (Politico earlier; confirmed circulating). Would allow a judge to halt release/use only for imminent catastrophic risk (>50 deaths or >$1 B damage) tied to bio, loss-of-control, or cyber. Lab-authored proposal; medium confidence on legislative prospects. Attempts to set a high bar for intervention.

10. **Moonshot states it removed hacking-related material from Kimi K3.1 Pro training**; China proposes broader AI export controls including on certain models/agents. Company statement + draft rules; medium confidence on effectiveness. Affects open-model cyber capability distribution and export/sovereignty dynamics.

(Additional notable: AI-assisted proofs or community revision of math results including Erdős-related conjectures; Anthropic internal data on junior technical hiring mix; various US state AI labeling/safety moves. Ranked lower on immediate Frontier/cost/legality impact.)

### Frontier one-pager
**Current usable Frontier cluster (independent evals, not lab claims):**  
Claude Opus 5.5 (max/fallback) leads Artificial Analysis Intelligence Index v4.3.2 (~57.6–58) and Epoch Capabilities Index (~169 / top). Closely followed by Claude Sonnet 5.5, Muse Wonder Full (Meta), GPT-6 Astra / GPT-6.1 Sol, and Gemini 4 Argon (ECI ~166, closest Google approach in months). Kimi K3.1 Pro now sits immediately behind the top proprietary cluster on AA (55).  

**Default Efficient generalist:** Claude Sonnet 5.5 (or equivalent mid-effort Opus 5.5) — near-frontier on many agentic/coding/knowledge tasks at lower token price and higher speed than max Opus.  

**Default Budget with evidence:** Smaller/cheaper models (Claude Haiku 5.5 class, strong open-weight MoEs such as DeepSeek V4.1 Flash variants or GLM-5.3 class, or Kimi once weights land). Evidence from AA cost-per-task and open-weight lag of ~6–7 months on Epoch Cyber ECI.  

**Availability caveats:** Opus 5.5 and peers are generally available via APIs/cloud but subject to rate limits, regional restrictions, and capacity constraints (AWS sold-out signals). Kimi K3.1 Pro weights not yet public (promised ~4 Nov). No evidence this week of new Grok 4.7 or Gemini 3.5 Pro general availability beyond prior releases; do not treat unconfirmed names as shipped. Training-time internet cut at OpenAI may slow future capability jumps. Separate all lab self-reports from AA/Epoch independent harnesses.

### Economics
- Capacity: AWS largely pre-sold through 2027 (Jassy); broader data-center spend forecasts remain elevated. Memory, power, and large-block reservations are binding constraints.  
- Pricing / cost-per-task: Kimi K3.1 Pro ~$2 / $10 (33% cut vs prior) and ~$2.11 per AA Intelligence Index task — competitive with or better than many closed models at similar quality. Opus 5.5 remains higher ($4 / $20). No major broad price cuts this week.  
- Open vs closed: K3.1 Pro (once weights release) is projected to become the leading open-weight model on AA Index, shrinking the gap to ~3 points. Other open MoEs continue incremental gains.  
- Capital & sovereignty: OpenAI ~$30 B / $1.25 T talks (UAE-linked); DeepSeek reported large China round (~$15 B). China export-control drafts and US investment-restriction proposals tighten cross-border model/compute flows. UK biosecurity package includes lab credits from US labs.

### Compendium edit proposals (maps-of-intelligence)
- **Accept:** Update Frontier cluster to reflect AA Index v4.3.2 and Epoch ECI leadership by Opus 5.5, with Kimi K3.1 Pro as near-frontier open-weight (pending weights). Note agentic/cyber strength of K3.1 Pro and PHP2Shell as evidence of practical cyber-research capability. Record OpenAI’s offline-world training policy as a change in training regime.  
- **Watch:** Exact post-weights ranking of K3.1 Pro; whether AWS (and peer) capacity constraints materially raise effective cost or latency for frontier usage; outcome of OpenAI funding round and any linked compute deals.  
- **Reject (for now):** Treating any unreleased or non-independently-evaluated model as Frontier; equating lab Terminal-Bench claims with AA harness results; assuming Grok 4.7 or Gemini 3.5 Pro general availability without primary confirmation.

### Task Router atlas notes
- Terminal-Bench 4.0 and agentic suites (GDPval-AA, AutomationBench-AA, CyberGym-style) remain discriminative; K3.1 Pro sets or approaches highs in AA’s harness.  
- No major new harness scandals this week. PHP2Shell shows models can accelerate vulnerability discovery outside standard coding benchmarks.  
- Evals with high saturation risk (easier knowledge or short-horizon coding) continue to compress; prefer long-horizon agentic, cyber, and science suites (Epoch, AA, FrontierMath-style) for ranking.  
- Open-weight cyber lag remains ~6–7 months on Epoch Cyber ECI (top open still behind closed frontier).

### Look-ahead (next 7 days)
- Expected Kimi K3.1 Pro full weights release window (~4 Nov).  
- Possible OpenAI funding-round signatures or updates (targeted end of prior week / early next).  
- Amazon Q3 results (29 Oct) — may clarify capex and capacity commentary.  
- Continued PHP2Shell patching and exploitation monitoring.  
- No other strongly evidenced major model or regulatory deadlines in the immediate window.

### Sources
Primary/independent: OpenAI blog/statement on training isolation; Artificial Analysis (Kimi K3.1 Pro article and Index); Epoch AI (ECI updates, Brief 23 Oct); PHP project / Wiz / Calif disclosures and Wikipedia/operational summaries; UK government AI Biosecurity Lab announcement; Reuters on AWS capacity (Jassy); CNBC/Bloomberg/Reuters reporting on OpenAI financing; EU Official Journal on AI Act standard; Anthropic product pages and Duet announcements.

Users may paste any task for a Task Router ticket.