1 / 11
← → navigate · scroll / swipe · F fullscreen
Episode 37 · Friday, November 6, 2026 · 4:00 PM ET

Weekly Claw

OpenAI rebuilt Git for its agents and then bought the tooling around it, Meta bought the multi-agent workspace, Kimi shipped a 2.8T checkpoint you can actually download, two practitioners fought over what “reading weights” means, and a $1.2M model run produced 948 working root exploits. The layer agents run on got consolidated, priced and attacked in five days.

Hosts: @AndyML · @HiM ~34 min · 8 stories · one video anchor
Sponsored by Herald Labs Heritage Telecom
“The agents got a landlord, a lawyer, and a bug bounty.”
Weekly Claw #37
01 / COLD OPEN
Cold open · the frame

The agent stack got a landlord. And a bug bounty.

Five days, one throughline: the layer agents run on is being consolidated, priced and attacked. Infrastructure broke under agent load, two labs bought the tooling around it, and the security receipts arrived in the same week.

1
The code stack
2
Open weights
3
Weight space
4
Agent security
5
Where agents run
🛠OpenAI’s Git backend went to ~1PB/day and reliability fell below 90% under agent concurrency. Two days later it bought Pierre; Meta bought Conductor the same day.
🧱Kimi K3.1 Pro posted its full weights — 2.8T total, 104B active, 1M context — ungated. ERNIE-5.5-Flash-Base landed the next day at 280B/14B-active.
🧠Qwen 3.8 27B named 3 of 4 games from raw policy weights on a laptop. Honnibal says the viral framing is misleading. Both are true.
🔒A $1.2M model run produced 948 working Linux privilege-escalation exploits — each with a patch. Check Point found the first prompt adware inside a Codex jailbreak.
💵Anthropic’s S-1/A: $16.96B Q3 revenue against $895.4B of purchase commitments. The revenue is the headline. The commitments are the story.
Weekly Claw #37
WHAT HAPPENED THIS WEEK · 1 / 4
What happened this week · the agent-era code stack

OpenAI rebuilds Git for agents, then buys the tooling around it.

OpenAI engineering post: Scaling Git at OpenAI, November 2, 2026 — building agentic-scale source control while investing in the Git ecosystemPrimary post · Nov 2

OpenAI ships Pando, a Rust Git backend built for agent-scale monorepos.

~1PB/day, median push latency cut ~5x, reliability had fallen below 90% under Codex concurrency. Built with Codex in 10 weeks, unmodified Git protocol.

▶ LIVE · openai.com · HN thread
OpenAI acquisition announcement: craftsmanship meets scale for the agentic era — OpenAI to acquire Pierre Computer CompanyPrimary post · Nov 4

OpenAI acquires Pierre, Meta acquires Conductor — same day.

Pierre built Code Storage and the Trees/Diffs libraries Codex already ships. Conductor’s multi-agent workspace cloud shuts down Nov 20.

▶ LIVE · openai.com · TechCrunch
Weekly Claw #37
WHAT HAPPENED THIS WEEK · 2 / 4
What happened this week · open weights

The open-weights barrage: one frontier checkpoint actually shipped.

Kimi K3.1 Pro vendor scorecard: 2.8T total, 104B active, 1M context, Terminal-Bench 4.0 at 65.7Model card · Nov 4

Moonshot posts Kimi K3.1 Pro weights — 2.8T total, 104B active, ungated.

1M context, MXFP4 QAT, MoonViT-V2 vision, Gated DSA on 24 layers vs K3. vLLM day-0; Huawei serves it on Atlas A3 and Ascend 950.

▶ LIVE · huggingface.co · weights
Baidu ERNIE-5.5-Flash-Base release card: sparse MoE 280B total, 14B activated per token, 33.5T pretraining tokensPrimary release · Nov 5

Baidu open-sources ERNIE-5.5-Flash-Base, and Tencent’s Hy4 weights are still pending.

ERNIE: 280B total / 14B active, 33.5T tokens, 1M-token fine-tune context, custom licence. Hy4 is API-only — vision and computer use, weights not posted.

▶ LIVE · ernie.baidu.com · hy.tencent.ai
Weekly Claw #37
WHAT HAPPENED THIS WEEK · 3 / 4
What happened this week · weight space

Qwen 3.8 27B reads its own weights. The rebuttal is already out.

Simon Willison weblog post: Qwen 3.8 27B on my laptop can interpret model weightsPractitioner post · Nov 3

Willison replicates it locally: Qwen 3.8 27B names 3 of 4 games from raw policy weights.

His own caveat, on his own page: “some of the wider commentary… does not feel justified.”

▶ LIVE · simonwillison.net
Honnibal blog post: the Astra can read and write weights to improve Qwen post is misleading, November 3, 2026Rebuttal · Nov 3

Honnibal: only toy nets are read weight-only; the Qwen and Go results lean on Python programs.

The “Sparks” claim says Astra fixed Qwen3.5-0.8B MATH-500 7.9→48.1 by editing 0.016% of weights. Claim vs rebuttal — not settled.

▶ LIVE · honnibal.dev
Weekly Claw #37
WHAT HAPPENED THIS WEEK · 4 / 4
What happened this week · agent security

A $1.2M model run produced 948 working root exploits.

OpenAI security publication, November 4, 2026: A thousand fewer paths to root — 948 local privilege escalation vulnerabilities validated and patched at $1.2 million API costPrimary post · Nov 4

GPT-6 Cyber ran a 30-hour pipeline that validated 948 Linux local-privesc exploits, each with a patch.

2,308 candidate findings → 1,724 reproduced → 948 working exploits at $1.2M of API. Launched Nov 5: ExploitGym 48.5% vs 32.1% for GPT-5.6 Cyber.

▶ LIVE · openai.com · THN
Check Point Research figure 1: contents of the Haiou jailbreak bundle package — prompt modules, skills, payloads and platform launchersResearch figure · Nov 3

Check Point finds prompt adware inside a Codex jailbreak bundle.

The Haiou bundle installs Skills and rewrites Codex config so ads for the author’s paid LLM gateway land in later agent responses. Seen beside Cobalt Strike and VulnClaw.

▶ LIVE · research.checkpoint.com
Weekly Claw #37
05 / SIGNAL FROM OUTSIDE · 6 MIN
Signal From Outside · weekly video review · permanent anchor · 6 min

Factory’s compute layer: agents shouldn’t care where they run.

THIS WEEK’S SIGNAL · NIKHIL HARITHAS · FACTORY · AI ENGINEER

“Your AI Agents Shouldn’t Care Where They Run”

Droid rebuilt so agents run inside a customer’s own cloud instead of borrowing a developer’s identity. The shape: a control plane, a compute gateway and a relay; bring-your-own-key model calls that stay in the VPC; snapshots and forking for fleets of agents; and a plan to open source the compute layer because the abstractions have nothing to do with Factory.

Verified: YouTube metadata, 2026-11-03 · 17:13 · 5,085 views at verification · transcript saved and cues read against it · spot-check before air.

Caveats on air: vendor talk — Factory describing Factory; conference talk, not an independent audit. Cue timestamps come from auto-captions — re-locate on open. Manual open only — NOT autoplay.

▶ WATCH ON YOUTUBE · 17:13 · manual open · NOT autoplay
Nikhil Harithas, Factory: Your AI Agents Shouldn’t Care Where They Run — AI Engineer talk poster

Cue beats (manual open, transcript-checked): 2:44 agents can’t keep borrowing your identity · 5:42 control plane, compute gateway and relay · 9:34 bring your own key, stays in the VPC · 11:21 open-sourcing the compute layer · 12:24 snapshots and forking · 14:22 the hard parts (8-hour microVM caps, rehydrate) · 16:44 “agents shouldn’t care where they run”.

Fallback still: this poster. If the embed will not play, the talk track stands on the cue beats above.

Weekly Claw #37
06 / HOT TAKE · 4 MIN
Hot take · the receipts

Anthropic’s revenue is the headline. The commitments are the story.

Anthropic quarterly revenue table, unaudited: $376M Q1 2025 rising to $16,956M Q3 2026, plus 787, 1,382, 4,731 and 11,555 interim bars

S-1/A · SEC EDGAR · unaudited, vendor-filed · Q3-26 $16,956M vs $895.4B of purchase commitments, ~13x annualized revenue.

The proposition

Nobody signs $895.4B of compute commitments by accident. That is a demand forecast — and if it is even 20% wrong, the balance sheet that absorbs it is not Anthropic’s.

What would change Henry’s mind: a disclosed commitment schedule where the $895.4B steps up only after revenue clears a stated threshold.

Henry

Thirteen times annualized revenue, signed before the roadshow. If compute demand flattens, the people holding the leases find out first.

Andy

Commitments are how you get priority allocation in a supply-constrained market. The question is not the size, it is whether the revenue curve keeps bending the way Q3 did — +46.7% quarter on quarter, $33.2B across nine months.

Weekly Claw #37
07 / CLOSE
One to watch + where to follow

See you next Friday.

Conductor’s shutdown

The multi-agent workspace cloud goes dark November 20. Where those teams land — and whether Meta ships anything that replaces it — is the first real test of the devtools land grab.

Anthropic’s roadshow

Nasdaq “ANT”, roadshow reportedly around November 10. Watch whether the $895.4B of commitments survives IPO diligence or gets re-cut before pricing.

Kimi in the wild

2.8T of ungated weights is only news if someone runs it. Watch for independent serving numbers — and whether Ripple-65B’s $32M training run turns out to be the better economics story.

The agents got a landlord, a lawyer, and a bug bounty.

Weekly Claw #37
SOURCES / LINKS
Every claim, one click · verified 2026-11-05

Sources & Links.

Dated captures · Anthropic figures vendor-filed · exploit counts OpenAI-reported · Hy4 weights pending; ERNIE licence is custom, not Apache · weight-space claims contested · SFO cues auto-caption timestamps, spot-check before air · Pando client-feature caveat per sxz.io.