· 13 min read
OpenAI's cyber agents patched 948 Linux root exploits for $1.2M; Claude found a Xen hole
Published · Updated
OpenAI's agent pipeline found and patched 948 Linux local privilege-escalation paths for about $1.2M, and Claude surfaced Xen bug XSA-523. OpenAI also showed Pando, a Git backend built for Codex-scale pushes, and Ripple-65B and Tencent Hy4 moved open models forward.
Today in 30 seconds
- OpenAI's cyber agents built 948 working Linux root exploits plus patches in one run, for about $1,279 each.
- Xen advisory XSA-523, found by Anthropic using Claude, lets a malicious guest read and write arbitrary memory, including Dom0.
- OpenAI's Pando Git backend cut median push latency by about 80% under Codex-scale commit volume.
- Ricursive's dense, Apache 2.0 Ripple-65B Preview claims 80.52 on Terminal-Bench 2.1, built for $32M.
- Menlo says enterprise AI deployment spend reached $176.8B, and agents generate about 80% of OpenRouter tokens.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
OpenAI’s cyber agents found, exploited and patched 948 Linux local privilege-escalation paths in one mid-October run, for about $1.2M in API credits. Anthropic’s Claude independently surfaced a Xen hypervisor arbitrary read/write bug (XSA-523).
On the build side, OpenAI published Pando, a Git backend that cut median push latency by about 80%, and Together AI turned inference optimisation into a continuous agent loop. Open models moved too, with Ricursive’s Ripple-65B and Tencent’s Hy4.
Top news
OpenAI’s cyber agents found and patched 948 Linux local privilege-escalation paths for $1.2M
OpenAI’s Patch the Planet pipeline, run with an upcoming cyber model in mid-October, produced 948 working Linux local privilege-escalation (LPE) exploits plus candidate patches. The end-to-end API cost was $1,213,430, about $1,279 per LPE.
- Pipeline stages:
- Scout agents selected about 5,034 audit targets (about 2.8M lines of code, roughly 7% of the kernel).
- Analyst agents produced 2,308 candidate findings; dynamic validation reproduced 1,724 (about 74.7%).
- ExploitDev agents built 291 info-leak PoCs and 1,169 working LPEs; variant analysis yielded 948 distinct patches.
- Realistic targets: the VMs used stock desktop protections (SMEP, SMAP, KASLR, HARDENED_USERCOPY), not a weakened lab kernel.
- Historical Pwn2Own / TyphoonPWN LPE bounties ($20k–$70k) imply a hypothetical value of $18.9M–$66.3M, 16–55x the run cost.
- Hardened kernel: in a kernelCTF-style environment the pipeline got 21 LPEs and 25 info-leak PoCs for about $32,936 (about 45x below a ~$1.5M bounty estimate). Agents used at least 10 novel techniques (details withheld).
- Already landing: during the 10-day scan, maintainers patched or staged 104 of the LPEs and were discussing about 300 more.
- A July GPT‑5.6 Sol run found 166 LPEs; the newer model found more than 5x as many despite three months of hardening.
- Early bug: 20 of 200 “successful” exploits in one run had rooted via a different 0-day than the assigned finding, so OpenAI tightened path verification. Agents also found a QEMU TCG LPE in the test harness itself.
Why it matters: Unpatched local accounts on out-of-date kernels are effectively root for a well-resourced attacker. Automated discovery is now cheaper than historical bounties; the bottleneck is validation, prioritisation and patch quality.
- Do this now: patch Linux aggressively and assume low-privilege shells need sandboxes (Firecracker, gVisor). If you run vulnerability agents, verify the exploit path, not just “got root”.
Sources: OpenAI — A thousand fewer paths to root · The Hacker News
Get this brief in your inbox every morning — free
Xen XSA-523: Claude found an arbitrary guest read/write via grant copy
Xen Security Advisory XSA-523 (CVE-2026-102703, CVE-2026-102704) lets a malicious guest read and write arbitrary memory, including Dom0. It was discovered by Anthropic using Claude, validated by Ada Logics and independently reported by CodeAnt.
- The bug: grant-copy batches on adjacent ranges cache validated state for performance, and the re-validation conditions were wrong in two ways.
- Frame numbers were mixed up with grant references.
- A check that frame-number copies are only allowed on the caller’s own memory was misplaced.
- Affected: all Xen versions from 4.6 onward; 4.5 and earlier are not. There is no known mitigation besides the patches.
- Released: 3 November 2026, with patches for xen-unstable / 4.19.x and 4.18.x–4.17.x.
Why it matters: Two frontier labs’ cyber agents hit serious bugs in core infrastructure in the same week. Agentic vulnerability research is no longer a one-lab novelty.
- Do this now: if you run Xen 4.6 or later, apply the XSA-523 patches immediately.
Sources: Xen XSA-523
OpenAI’s Pando: a Git backend for Codex-scale monorepo pushes
OpenAI rebuilt its monorepo Git backend as Pando, cutting median push latency by about 80% while serving nearly a petabyte of Git data per day. One engineer designed it, Codex agents implemented it largely in Rust, and it reached production in 10 weeks.
- Design: Git-protocol compatible, with causal consistency per branch and reusable packfiles in blob storage.
- Why it was needed: Codex drove a surge of parallel branches and pushes, and the old system’s reliability dipped below 90%.
- Pando drops repository-wide consistency so unrelated branches don’t contend.
- It reuses packfiles instead of generating one per client, with over 99% cache hit rates via regional Express caches.
- Small packs: on push, Pando advertises a selective trunk base so packs stay reusable.
- Median pack: under 20 KiB.
- p99 pack: under 3 MiB.
- Compaction mirrors LSM-tree levels (L0–L4).
- Upstream and verification: OpenAI contributed
packfile-uri, parallel pack download/index and “explicit haves” improvements to Git. GPT‑6 Astra agents wrote Lean/TLA+ formalisations and red-teamed the system for hundreds of hours. - Availability: OpenAI is exploring a wider release and already ships its Git build to developer machines and the ChatGPT desktop app.
Why it matters: Agent fleets break classic source-control assumptions. If your org runs many coding agents, Git coordination, not model quality, may be the next bottleneck.
Sources: OpenAI — Scaling Git at OpenAI
Together Compiler: agents continuously retune production inference
Together AI launched Polymorphic Inference and Together Compiler on 4 November: production loops where agent swarms rewrite and verify GPU inference code as models, traffic and chips change. It is already live on popular serverless endpoints.
- Motivation: 80% of OpenRouter traffic hits models released in the last three months.
- Architectures, traffic shapes and hardware (B200 → GB300 → Vera Rubin) shift too fast for static vLLM/SGLang configs.
- How it works: the Compiler treats optimisation as agentic search over persistent high-fidelity engines. Continuous Compilation adapts without hand-written kernels per workload.
- Access: limited beta for customer dedicated endpoints.
- Roadmap: shape-shifting speculative decoding, ThunderAgent Router, Together Post-Training and TogetherLink model routing.
- Claim: Together says it served more OpenRouter tokens than closed labs in October.
Why it matters: Inference is becoming a living system. Static “best config” posts age in weeks; the durable advantage is a closed loop that re-optimises against real traffic.
Sources: Together AI — Polymorphic Inference
Ripple-65B Preview: US open-weight coding SOTA for $32M
Palo Alto lab Ricursive Intelligence released Ripple-65B Preview, a dense 65B Apache 2.0 model it says beats much larger US open models at agentic coding. The team traces its lineage to Azalia Mirhoseini and Anna Goldie.
- Specs: 12T pretraining tokens, up to 1M context.
- Claimed scores:
- Terminal-Bench 2.1: 80.52
- SWE-Bench Pro v2.0 Hard: 79.1
- Cost: $32M full R&D from scratch, about one-tenth of prior open-weight coding SOTA compute commitments (their estimate).
- The RL climb used only 1.4M rollouts, about 70x fewer than Reflection Beam.
- Architecture: a dense thin-and-deep Transformer (d_model 5120, 80 layers) co-designed for NVIDIA training and serving, at about 40% MFU without custom kernels.
- It runs on consumer-class boxes such as DGX Spark; the community is already clocking NVFP4 throughput on RTX 6000 Pro.
- Where: weights on Hugging Face; hosted on Together, GMI Cloud and OpenRouter (Fireworks the same day). An internal variant powers Ricursive’s AI chip-design platform.
Why it matters: Dense, carefully laddered mid-size models can lead agentic coding without MoE mega-parameter counts. US open weights still have a path without Colossus-scale budgets.
Sources: Ricursive — Ripple-65B · Hugging Face
Tencent Hy4 adds vision and computer use; open weights coming
Tencent released Hy4 on 4 November with first-party vision and computer use (CLI and GUI), plus specialised training on Blender and CAD. Weights and system cards are planned in the coming days.
- Pricing: same API pricing as the Preview.
- Input: about $0.90 per million tokens.
- Output: about $2.70 per million tokens.
- Finance eval: a blind test with 213 professionals on 579 tasks.
- Hy4: 3.04
- DeepSeek-V4.1-Flash-1017: 2.97
- Kimi K3.1 Pro and GLM-5.3: 2.68
- Hallucination rate: 7.2% (Preview: 12.95%).
- Games: GameDev up 19.1% and GameLogic up 9.8% vs the Preview, via Unreal/Unity/Godot integration. A TenureTrail science suite covers 123 tasks across seven disciplines.
- Safety framing: trained on defensive cyber PoC and remediation data, with offensive exploitation material excluded.
- Xuanwu / Keen labs found “no significant concerns” on offensive capability in tested areas (crypto reverse engineering, deobfuscation, exploitation).
- Biosafety: strong knowledge, limited open-ended experimental design.
- Also: HyCell (single-cell) and HyPDE (physics fields) open-sourced via the Hyra 2.0 harness. Available on TokenHub, WorkBuddy, CodeBuddy, OpenRouter, Hermes and OpenCode.
Why it matters: Another Chinese frontier open-weight push with real computer use, as Menlo shows Chinese models already take 89% of open-source spend (item 8).
Sources: Tencent — Hy4
A heart means recursion: simple images steer VLMs without them noticing
Google DeepMind and Stanford (Alexa Tartaglini et al.) show a red heart image can flip Claude Opus 4.7 and GPT-5.4 from 0% to 100% recursive solutions on binary-tree traversal. The models’ thinking traces usually dismiss the image as irrelevant.
- Model-specific symbols: across a 288-symbol library (24 shapes × 12 colours), hearts and triangles steer Claude and GPT-5.4; a pink plus or cyan spiral steers Gemini 3.1 Pro.
- Matching text captions barely steer; Gaussian noise and ImageNet photos do not.
- Unfaithful traces: when symbols steer, about 71.5% of Gemini traces disregard them. Silent injection (no “look at the image” instruction) still steers, often with different symbols.
- Sender–receiver games: models can communicate concepts via free-form PIL drawings (about 91% zero-shot).
- Symbol generation and interpretation diverge, and multi-turn symbol communication becomes less human-legible as it succeeds.
- Agent risk: agents that see screenshots or GUIs can be steered unintentionally, in ways chain-of-thought oversight will miss.
Why it matters: If your agent looks at a screen, “ignore the decoration” is not a reliable instruction. Visual inputs are a control channel, not just context.
Sources: arXiv:2610.36449
Menlo: enterprise AI spend hit $177B, and agents make 80% of tokens
Menlo Ventures’ 2026 State of Enterprise AI puts deployment spend at $176.8B by September 2026, up 5.4x from $33B. On OpenRouter, agents generate about 80% of tokens, about 7x humans.
- Survey: about 400 US decision-makers.
- Who gets paid: frontier labs take 69% of deployment dollars (up from 50%).
- Enterprise LLM API spend: Anthropic 55%, OpenAI 31%, Google 9%.
- Open source: 19% of tokens but only about 4% of spend.
- Chinese models take 89% of that open-source spend: Kimi 61%, DeepSeek 15%, GLM 11%.
- Apps: a $77.3B market; coding ($34.0B) and chatbots ($25.0B) make up 76%.
- Claude Code + Codex: $24B, 70% of coding. Coding is 96% AI-native.
- Budgets: average enterprises spent 4.7x 2025 levels vs a planned 1.9x.
- 55% of incremental AI dollars are net-new budgets, not software or labour cannibalisation.
- Every $1 into model development yields about $1.70 in deployment revenue.
Why it matters: The “SaaSpocalypse” framing looks wrong in Menlo’s numbers. AI is expanding software spend, and agents, not seats, drive demand.
Meta buys Conductor; Google backstops TeraWulf’s $19B Anthropic leases
Meta is acquiring multi-agent coding startup Conductor, whose cloud shuts on 20 November (the local app runs until early 2027). Separately, TeraWulf got Google credit support for financing about $19B of Anthropic leases at Hawesville, Kentucky.
- Conductor: founded by Charlie Holtz, it runs parallel coding agents in isolated workspaces on top of Anthropic and OpenAI models.
- Holtz says he’ll work with Alexandr Wang and Nat Friedman at Meta Superintelligence Labs. Meta confirmed the deal, no terms.
- He teases a new interface early next year under Muse.
- TeraWulf: Google LLC credit support covers project debt tied to Anthropic PBC leases; Alphabet is rated AA+/Aa2.
- Fluor moves to a full GMP-basis EPC for three buildings (ready for service late 2027–Q2 2028).
- TeraWulf builds power-backed infrastructure, not the GPUs themselves.
Why it matters: Meta keeps buying agent UX talent for Muse Code. Anthropic’s compute commitments are large enough that Google will underwrite the landlord’s financing.
Sources: Conductor announcement · TechCrunch · TeraWulf press release
GPT-Live outbound telephony hits GA; India is ChatGPT’s largest market
OpenAI made GPT-Live outbound telephony generally available, with SIP trunks (including Twilio and Telnyx), HD Voice and DTMF. Separately, India has overtaken the US as ChatGPT’s largest market by weekly active users.
- How to call out:
POST /v1/live/sessionswithtransport.type: "sip"and an E.164 destination. It builds on the GPT-Live-1 voice API; juberti flagged the GA on X (5 November). - India: Altman and OpenAI India MD Prabhjeet Singh both confirmed the shift, after India’s user base roughly doubled since January.
- ChatGPT overall:
- About 1.2B weekly active users, after +200M in two months.
- About 250M weekly voice users.
Why it matters: Voice agents can now dial out as a first-class API surface. ChatGPT’s growth centre has moved east, and product, pricing and policy will follow the users.
Sources: OpenAI — GPT-Live-1 in the API · Economic Times · @juberti
GitHub repos
- View repo →
NVIDIA/SkillSpector
A security scanner for AI agent skills that looks for malicious patterns, prompt injection and data exfiltration. It fits the week’s agent supply-chain theme.
- View repo →
max-sixty/worktrunk
A CLI for Git worktree management aimed at parallel AI agent workflows. A natural companion to OpenAI’s Pando post if you run many agent branches locally.
- View repo →
Anil-matcha/open-duet
A self-hosted open alternative to AI coworkers such as Grok Bot, Claude Duet, OpenAI Dots and Meta Muse.
Research papers
- Read paper →
A Heart Means Recursion
Simple images steer VLMs while their reasoning dismisses them (item 7).
- Read paper →
Route Anything
Model and policy routing for general-purpose AI systems.
- Read paper →
PyRUA-DataFarm
An LLM agent as a cost-efficient teleoperator for robot data.
- Read paper →
From Pixels to Actions
Native vision-language-action policies.
- Read paper →
SpecOPD
Branch-parallel on-policy distillation for speculative decoding.
AI workflows
1. A 50-person company where agents watch the funnel every morning and fix bugs from Sentry · Workflow page
- Replaces: a human PM reviewing the product funnel once a week, manual competitor and review research, and hand-triaging production errors.
- At 7am an agent reads yesterday’s PostHog funnel; for any step dropping more than 10%, it watches 20 session replays and explains why.
- Sentry errors get a fix PR in under 10 minutes that a human only approves.
- Every Monday all Stripe cancellation reasons land in one doc.
- What to connect: Grok Bot (voice notes in), Claude Code, Linear MCP for specs, and Exa + Firecrawl and Apify for competitor sites and G2/Reddit/App Store reviews.
- PostHog, Sentry, Stripe, Supabase, GitHub and Vercel MCPs.
- A git-backed markdown memory and a nightly 5-line log per agent.
- Rules that make it work: one agent per platform, a cheap model for triage and a strong one for anything that ships, and every agent mistake turned into a new skill.
- Who it’s for: founders and small product teams running a SaaS who want a daily, not weekly, read on funnel drops, churn reasons and errors.
2. Explainer cartoons from one Claude Code folder: first video 3 hours, fifth 25 minutes · Workflow page
- Replaces: drawing and animating short explainer videos by hand or paying an animator.
- Results (claimed): in the first week, 5 videos of 35–50 seconds, averaging 40 minutes each with edits, for $0 in software plus a $30 Claude subscription.
- What to connect: Claude Code in local mode inside one folder with a guide.md of style rules.
- An SVG character with a rig and 8 mouth shapes; scenes as numbered React components; Remotion for the timeline.
- Whisper for word timestamps and captions, Rhubarb Lip Sync, and ffmpeg for the render and loudness.
- How it works: edits go by scene number (“3, camera closer”), and after the third clean video the whole process is saved as a skill.
- Who it’s for: creators, educators and brands that want a recurring explainer series on Reels, Shorts or TikTok without an animation team.
3. Shipping rules fixed and a weekly sales report from a store owner’s bot · Workflow page
- Replaces: hours of manual work rewriting complex shipping policies in an online store and analysing sales by hand.
- The bot made the changes itself and now sends a weekly report with performance metrics and revenue recommendations. The author says setup took about 10 minutes.
- What to connect: Grok Bot with access to the store admin and its sales data, plus a weekly routine for the report. No cost stated.
- Who it’s for: small e-commerce owners who run their own store and lack time for weekly analysis.
Quick hits
- xAI turns on invisible watermarking for Grok text — from 7 November 2026, grok-4.7 / 4.6 / 4.5 / 4.3 / 4.20-* responses get a statistical token-level watermark by default, for the EU AI Act. xAI docs
- Check the
x-watermarkheader; team admins can opt out per model in Console. No request-parameter change.
- Check the
- Baseten ships Grid, Sandboxes and Loops GA — model routing and governance, resumable VM-isolated agent sandboxes, and serverless SFT/RL loops where customers keep trained weights. (via TLDR AI, 4 Nov)
- DeepSeek plans to pretrain V5 on Huawei Ascend — the first serious test of Ascend for frontier pretraining, not just inference. (via TLDR AI, 4 Nov)
- If it works, Chinese labs depend less on Nvidia.
- ChatGPT Images will make GIFs — animated GIF output has been merged into OpenAI’s Images API path and is coming to ChatGPT. (via TLDR AI, 4 Nov)
- Artificial Analysis makes time-per-task its headline metric — it measures full request latency across about 1,000 serverless endpoints. Artificial Analysis
- GPT-6 Astra leads max-effort agentic tasks in circulating summaries: about 17 min vs about 26 for Opus 5.5 and about 29 for Gemini 4 Carbon.
- Default medium effort cuts time 3–5x.
Go deeper
- OpenAI: Scaling Git at OpenAI — the full Pando post, with the consistency-model comparison and simulation.
- OpenAI: A thousand fewer paths to root — the cyber pipeline methodology (Scout → Analyst → Validate → ExploitDev → Patch).
- Menlo Ventures full report (PDF) — all the enterprise AI spend numbers.