· 17 min read
Gemini 4 Carbon tops the new AA index, agents breach Anthropic via SaaS, Alpine's GitLab exposed
Published
Google's Gemini 4 Carbon leads the rebuilt Artificial Analysis index but goes to cyber defenders first, AI agents got into Anthropic through third-party business software, and Alpine's patched GitLab was still exposed.
Today in 30 seconds
- Google's Gemini 4 Carbon scores 59 on the new Artificial Analysis Intelligence Index v5, but only cyber defenders get access this week.
- Cathedral's AI agents reached some of Anthropic's internal sales, HR and IT systems through third-party tools during a bug bounty; the two sides dispute the details.
- Alpine Linux rebuilt its GitLab after an attack on CVE-2026-97229, because upgrading GitLab alone didn't fix the Ruby bug in a self-built image.
- A fingerprinting script hides jailbreak bait so LLMs refuse to deobfuscate it, and a RISC-V CPU now runs inside Qwen 3.5's weights.
- GitHub's Octoverse shows agents bent the curve in June, and Microsoft booked $10B of revenue from OpenAI in one quarter.
Free. Daily, 10-min read. Confirm by email, unsubscribe in one click. Privacy
Pour the tea and settle in: Google shipped Gemini 4 Carbon only two weeks after Argon, and it goes straight to the top of the newly rebuilt Artificial Analysis index, but only cyber defenders get it this week. Security is the thread through the rest of the pot.
AI agents got into Anthropic through third-party office software, Alpine Linux shows how a “patched” GitLab can still be exposed, and someone hid jailbreak bait inside a fingerprinting script to scare off LLM-assisted reverse engineering. For the tinkerers there’s also a RISC-V CPU compiled into Qwen 3.5’s weights.
Top news
Gemini 4 Carbon tops the new AA index, but goes to cyber defenders first
Google DeepMind announced Gemini 4 Carbon on 28 October, two weeks after Argon. It scores 59 on Artificial Analysis Intelligence Index v5 (Argon: 50) and sets new state-of-the-art results on cyber benchmarks.
Because of those cyber capabilities, it’s going through the US government’s voluntary pre-release testing and goes to a small group of cyber defenders first.
- Cyber benchmarks (run without guardrails):
- 88.2% on SEC-bench Pro (finding vulnerabilities).
- 54.8% on ExploitGym (turning them into exploits).
- 71.0% on Collinear’s CWE-bench (patching).
- Google says Carbon writes secure code more than 2× as often as Argon in internal tests. It also says the Chrome team found it more than 10× better at writing patches than “another lab’s large frontier model”.
- Long-horizon agent work: Google says internal teams ran Carbon for days on root-causing and patching subtle bugs in large codebases and low-level systems. WebDev Arena rose 89 Elo to #4 (1,768).
- Alignment claims (Google’s numbers):
- Cheating on Terminal-Bench 4.0 fell to 0.0% (cheat attempts 0.4%).
- High-severity misalignment alerts in internal live coding dropped by almost 80%.
- Unsupported claims are about 2× lower than Argon. Overreach is down 26% and under-service down 29%.
- Gray Swan’s indirect prompt-injection attack success rate is 0.0% (Argon: 0.7%).
- Independent numbers (The Decoder roundup):
- Carbon trails Claude Code with Opus 5.5 by three points on AA’s Coding Agent Index (59 vs 62).
- AA-Omniscience accuracy rose from 48% to 58%, but the hallucination rate also rose, from 4% to 11%.
- It places second in the Vals Index (70.36%), just behind Sonnet 5.5.
- It’s second in Arena’s Agent Arena, with a $1.80 median task cost, 72% below Opus 5.5 (max). On AA it costs $7.76 per task, about the same as Argon.
- Rollout and price: a subset of cyber defenders gets access this week. Access then widens to trusted partners through the Fairwind Program, and developers and consumers get it later.
- Introductory pricing matches Argon at $2/$10 per 1M input/output tokens (cached input 95% off), rising to $4/$20 after the intro period.
Why it matters: labs are now shipping new frontier models every two weeks, and the newest ones are deliberately held back because of their cyber skills. Expect “available to defenders first” to become the normal release pattern for top models.
Sources: Google blog · The Decoder benchmark roundup
Get this brief in your inbox every morning — free
AI agents broke into Anthropic through third-party business software
The security startup Cathedral, founded by former DOGE staff, used AI agents during Anthropic’s bug bounty to get into some of Anthropic’s internal sales, HR and IT systems through external admin and sales tools.
The two companies disagree on how long it took and how far it reached.
- Cathedral’s draft report, reviewed by The Washington Post, says its agents reached meeting transcripts, salary data, job postings and Slack messages in under 12 hours.
- Anthropic confirms access to a limited set of sales, HR and IT systems, including some customer and partner contact details and basic account data.
- It says Claude chats, source code and model weights were never exposed.
- It shut the access points the same day, disabled the third-party tools and notified US federal authorities.
- Anthropic red teamer Logan Graham says the probing lasted several days. He says Cathedral found two vulnerabilities but reported only one, which Anthropic says broke the bounty’s disclosure rules and went beyond its safe harbor.
- Former OpenAI researcher Will Depue says his sources describe full admin access to Slack. Graham says those sources are wrong but agrees the industry “has to reinvent its security every year”.
- Graham’s advice is to red-team yourself constantly and rebuild software with formal verification.
- He says even models half a generation behind the frontier can run these attacks around the clock.
Why it matters: the weak spot wasn’t the model. It was the SaaS sprawl around it. Anthropic arguably has the best AI cyberdefense tools in the world and still got hit through its integrations.
Sources: The Decoder (cites The Washington Post, Graham and Depue)
Artificial Analysis Index v5 moves to private tests, and the open-model gap grows to 12 points
Artificial Analysis rebuilt its Intelligence Index. 60% of v5 now uses private test sets or held-out answers.
The best open model now trails the leader by 12 points, up from 3.
- New in v5: Cognition’s private FrontierCode 2.0, which grades whether a maintainer would merge the agent’s change.
- Also new: Terminal-Bench-Science (70 agentic research tasks), a refined 1,000-question HLE-Diamond, and AA-Omniscience v1.1.
- Before Carbon arrived, Claude Opus 5.5 (max) led overall. No model leads all nine evals. On the hard CritPt subset GPT-6 Astra scores 72% and Opus 5.5 61%.
- Kimi K3.1 Pro (weights not yet released) is the best open model at 46.
- Among models with released weights, DeepSeek V4.1 Flash 1017 and GLM-5.3 lead at 32, and Claude Haiku 5.5 beats both.
- Grok 4.7 lost 12 points in the switch.
- Cost: GPT-6.1 Sol (max) lands within 4 points of Opus 5.5 at about one-eighth of the cost per task ($1.27 vs $10.04). Median output tokens per task rose from 35k to 59k.
- Nathan Lambert (Interconnects) says the index’s attempt to “reflect real-world usage” runs into jagged capabilities. Coding is where open models keep up best.
Why it matters: when a big part of a leaderboard is private, models that were tuned to the public test sets fall. Read the v5 numbers as a reset, not a trend line.
Sources: Artificial Analysis v5 announcement · The Decoder · Interconnects
Alpine’s GitLab shows how upgrading GitLab can leave you exposed to CVE-2026-97229
An attacker went after Alpine Linux’s GitLab on 20–21 October using CVE-2026-97229 (CVSS 10.0, a regex bug in Ruby’s Onigmo engine reached through CI config).
Alpine found no proof it succeeded but couldn’t rule it out, so it rebuilt the instance from scratch. The bigger lesson: upgrading GitLab didn’t fix the bug for them.
- A newly registered account sent crafted CI configs that crashed worker processes 8 times. In 4 of those crashes the exploit got partway through.
- The attacker also downloaded Alpine’s GitLab container image, likely to tune the exploit to their exact build.
- Alpine was on 19.3.4 by mistake instead of 19.3.5, which had the fix. Its later upgrade to 19.3.6 still left it vulnerable.
- That’s because the fix is a patch to Ruby that GitLab applies only in its official packages, and Alpine’s image used standard Ruby.
- Every password was reset and GitLab’s secret keys were replaced.
- The aports repository, packages built 20–24 October and the 21 October releases were checked against records outside GitLab and look clean.
- Package signing keys were never on that server.
- Alpine also warns that a successful exploit would leave no crash and nothing in the web logs.
Why it matters: any self-built GitLab image (built from source or with your own Ruby) is still exposed after a GitLab upgrade, and attackers are already tuning exploits to specific builds.
Sources: Alpine Linux incident post
A fingerprinting script stuffed with jailbreak bait to stop LLM-assisted deobfuscation
A researcher crawling canvas fingerprinting on 10,000 sites found an anti-bot script whose variable names, invisible Unicode text and dead code look like bioweapon, self-harm and “UNRESTRICTED mode” jailbreak prompts.
The goal appears to be getting models to refuse when asked to deobfuscate it.
- Strings like a “lethality threshold” summary sit next to an invisible-Unicode prompt asking for DDoS, bioweapon and drug guides.
- Shell commands echo fake
<thinking>self-harm text, and a block of Tiananmen and Taiwan terms looks aimed at Chinese models.
- Shell commands echo fake
- It works: GPT-6 Astra refused to deobfuscate it, even with the Daybreak Blue toggle on. GLM-5.3 complied.
- After the decoys are stripped out, it’s an ordinary anti-bot fingerprinting script, though it includes Node-only code paths behind an environment flag that fetch and beacon a “metrics-agent” binary.
Why it matters: this is a new way to fight back against AI-assisted reverse engineering: use the model’s own safety training as camouflage. Security teams that rely on frontier models for code analysis should expect refusals by design. They’ll need a fallback model or a tool that strips decoy text first.
Sources: nullpt.rs write-up
Qwenputer: a working RISC-V CPU compiled into Qwen 3.5’s weights
Divyansh Agrawal compiled a 32-bit RISC-V CPU directly into the weights of stock Qwen 3.5 4B, with no training. The model’s own Gated DeltaNet layers serve as exact, addressable RAM.
The CPU runs arithmetic, C and MicroPython inside the forward pass, with no external tool.
- Setting the DeltaNet gate to 1 and using one-hot keys turns each head’s state into RAM with no fading.
- The CPU (RV32IM plus Zba and a small bignum unit) takes 11 layers in front of Qwen’s 32.
- It passes the official riscv-tests and ran 18M instructions free-running without a single error.
- Qwen calls it through the normal tool-call format. NUPA long-integer arithmetic goes from 48.6% to 100%, the same as giving Qwen a Python tool.
- About 230k output positions were replayed against a reference machine with zero mismatches. Normal text output stays within noise of base Qwen.
- It also has a hashed in-state filesystem that persists between chats when you save and reload the state. It gets changed facts right 90% of the time, versus 30% when old chats are pasted back in.
- Limits: the checkpoint is 18.8 GB (mostly zero padding), it needs the RvChat driver, and it’s slow (about 3,200 CPU tokens/s batched on an L4). It runs on one 24 GB GPU.
Why it matters: besides being a great hack, it’s a safety testbed. The author’s follow-ups show hybrid RNN-attention models can hide reasoning in their recurrent state to evade chain-of-thought monitoring, and that compiled weights could serve as a monitor the model can’t turn off.
Sources: Blog post · GitHub repo · HF weights
GitHub Octoverse 2026: agents bent the curve upward in June
Developers pushed to GitHub 17.9B times in the past 12 months, more than in 2024 and 2025 combined.
Activity jumped around June, when agents went from experiments to persistent everyday tools.
- Commits across all repos reached 34.3B, over 3× two years ago. Merged PRs more than tripled to 1.3B (3.5M a day).
- Simon Willison, at GitHub Universe, reports 220M merged PRs and 7.38B commits per month now.
- New repos with an agent configuration file got 8× as many commits as new repos without one. From June on, about 55% of all commits came from brand-new repos.
- In September, 61% of Copilot users used more than one model. Actions ran 1,440 jobs a second, more than 3× last year. Over 30M Dependabot PRs were merged.
- New this year: industry breakdowns. Mining, construction, health care and agriculture repos have each grown several times over since 2022.
Why it matters: this is hard platform-wide data that agentic coding changed output this year, not just the demos. It also means far more code needs review, which is where the next bottleneck sits.
Sources: GitHub Octoverse 2026 · Simon Willison’s notes
Android Studio now runs any of 40+ coding agents, and Android apps become agent tools
At Android Dev Summit, Google added “Bring Your Own Agent” to Android Studio. Through the open Agent Client Protocol, more than 40 agents (including Claude Agent, Codex and Antigravity) can work directly with the IDE’s tools.
- Agents get the project structure, symbol resolution, the emulator and Logcat.
- A new Change Review tool explains pending AI edits and flags bugs, and a migration assistant converts XML layouts to Jetpack Compose.
- The Android CLI is now an official plugin for Claude Code, Codex and Antigravity.
- With screen preview in the terminal, agents finished UI tasks 77% faster with up to 3× fewer tokens.
- With LeakCanary built in, Opus 5.5 found and fixed memory leaks nearly 2× faster with about 25% fewer tokens.
- Google also ships 30+ skills for agents (Play policy checks and more).
- On the app side: an alpha Agentic API runs on Gemini Nano 4 on-device.
- An Android MCP beta lets apps share functions with system agents, and A2UI lets agents reply with the app’s own Compose UI.
- Over 50 partner integrations are live in the Gemini app.
- Hardware: XREAL AURA, the first wired Android XR glasses ($1,099, ships mid-November), with five virtual windows and Gemini that sees both the room and the screens.
Why it matters: Google is opening its flagship mobile IDE to competing agents and making apps callable by agents. That treats “which agent” as the developer’s choice and makes Android a big new surface for MCP.
Sources: The Decoder
Microsoft made $10B from OpenAI in one quarter
In its first fiscal 2027 quarter (to 30 September), Microsoft booked $10B in revenue from its commercial arrangements with OpenAI, nearly 11% of its $92.3B total.
It also disclosed Azure revenue in dollars for the first time: $33.2B.
- OpenAI revenue for all of fiscal 2026 was $24.1B. At the latest pace, that’s about $40B a year.
- OpenAI’s outstanding receivables to Microsoft rose to $7.2B from $6B in June.
- Microsoft’s stake fell to about 24% from 25% in June (about 27% after the October 2025 recapitalization).
- Under the April 2026 deal, OpenAI keeps sharing revenue with Microsoft through 2030, subject to a cap, while it’s free to use other clouds.
- TechCrunch hasn’t confirmed how much is revenue share versus Azure compute.
Why it matters: even after the partnership was loosened, OpenAI is one of Microsoft’s largest single revenue lines. Every OpenAI compute or pricing move now shows up in Microsoft’s results.
Sources: TechCrunch on OpenAI revenue · TechCrunch on Azure
Claude Code may spend 15–20% of your tokens on a three-word status line
A widely shared gist from Flamerix AI (134 points on HN) says Claude Code asks each background agent’s model (usually Opus) for a 3–5 word summary of its latest action about every 30 seconds.
The answer only feeds the agents panel’s status line.
- The authors measured roughly 15–20% of token use on this summarizer. The summary never reaches the model’s context or your conversation, and there’s no setting to turn it off.
- On a subscription, you don’t pay more for this. You lose rate-limit headroom.
- Their fix is a Bun preload hook (MIT license, binary not modified) that defuses only the summarizer’s timer. It was tested on Linux with Claude Code 2.1.257–2.1.312 and has an A/B control mode.
- This is a third-party measurement. Anthropic hasn’t responded in the sources we found.
Why it matters: if you run many background agents and hit limits, this overhead could be a real share of your usage. It’s also a reminder that agent harness features have token costs that don’t show up in your conversation.
Sources: Flamerix gist
GitHub repos
- View repo →
JetBrains/thinkrail
A lightweight agentic IDE from JetBrains built around the pi coding agent. Each session runs in an isolated git-worktree workspace, so agents work in parallel without touching your checkout. “Living specs” and reusable skills give the agent lasting project knowledge. It shows PR status, CI checks and review decisions, and can hand a failing check or diverged branch back to the agent. Full extension support is next. It’s notable because a big IDE vendor is shipping a small, agent-first editor alongside its heavyweight IDEs.
Research papers
- Read paper →
DTCA: Denoising Token Credit Assignment for RL in Diffusion LMs
Gives diffusion language models token-level RL credit based on confidence gains and entropy calibration, and improves math reasoning.
- Read paper →
Falcon: Efficient Streaming Audio-Visual Target Speaker Extraction
A lightweight causal audio-visual model that isolates one speaker in noisy real-world scenes, with a new RealSSA benchmark.
- Read paper →
OmniShowcase
Keeps a subject consistent in 3D across new viewpoints in video generation by training on geometry-driven data with surface-normal conditioning.
- Read paper →
TAO: Tool-Aware Joint Scheduling for Agentic Serving
Schedules agent sessions around tool readiness with dynamic programming, cutting latency and raising throughput.
- Read paper →
Sensorimotor Alignment
Offers a cheap offline predictor of robot policy success by measuring how well sensory representations match the geometry of physical motion.
AI agent use cases
- Check your OpenAI API credits: eligible ChatGPT Pro and Business subscribers can now claim up to $500 a month in API credits, so check your Pro account today.
- Patch your self-built GitLab: if you run GitLab with your own Ruby, upgrade Ruby to 3.3.13, 3.4.12, 4.0.9 or later in addition to GitLab, and re-pull any image older than 27 October.
- Audit third-party app access: list every third-party app with admin or API access to your workspace (Slack, CRM, HR tools) and revoke what you don’t need.
- Test the Claude Code summarizer hook carefully: if limits bite, run the hook in its A/B control mode first and compare your own usage before keeping it. Hooks that patch runtime behavior can break when Claude Code updates.
AI workflows
1. Customer support for 5,000 customers run by two Claude agents · Workflow page
- Replaces: a support person reading two inboxes and a feedback form by hand, with no shared queue and customers falling through the cracks.
- How it works: every message becomes a CRM ticket with six statuses, and a classifier filters out non-tickets and sets priority.
- A first-line Claude agent works the queue every 2 hours using 15 markdown knowledge files (billing, connectors, pricing, legal).
- A nightly auditor reopens threads where the customer wrote last, and a bug-triage agent checks Sentry and the code before engineering sees anything.
- A second-line agent posts one Telegram report every 90 minutes with numbered yes/no questions. Money, refunds and legal always wait for a human.
- Results (last 30 days): 15,500 emails in, median first reply 54 minutes, 72% under two hours, support headcount unchanged.
- What to connect: Claude, Attio (CRM), Gmail API with an hourly copy of all mail into a database (about 250,000 messages), Stripe, Sentry, the product database, the code repo and a Telegram group.
- Who it’s for: small SaaS and e-commerce teams whose support volume grows faster than the team.
2. Product photos to a layered, print-ready Photoshop file in about 15 minutes · Workflow page
- Replaces: days of manual design work on a product layout.
- How it works: the author generates 10–20 design directions from about 5 of their own studio product photos, picks the best, and runs ~5 more iterations of each style.
- Claude Code with a Photoshop MCP then rebuilds the chosen design as a real layered, high-resolution production file.
- Results: it took about 15 minutes while doing three other things. The author says it removes the dirty work but still needs adjustments.
- What to connect: an image model for the style iterations (Grok Bot, with Gemini as a fallback), Claude Code, a Photoshop MCP server and Photoshop, plus your own product photos.
- Who it’s for: small product businesses and in-house designers who make catalogs, packaging or promo layouts from studio photos.
3. A one-person local service business set up with Claude in three mornings · Workflow page
- Replaces: a marketing agency for a hull-cleaning diver who had no Google Business Profile, no reviews and no real website.
- How it works: in about 3 hours spread over three mornings, the owner learned to work with Claude like an employee.
- He set up an automatic review request after every job and a CRM that tracks every customer and follow-up.
- He also built a new website he now writes content for weekly.
- Results: two months later, 34 reviews and the #1 local ranking, all run by the owner himself.
- What to connect: Claude, a Google Business Profile, an automated review-request flow, a simple CRM and a website builder.
- Who it’s for: solo tradespeople and small local service companies with little or no online presence.
Quick hits
- OpenAI gives Pro and Business subscribers API credits — @OpenAIDevs (reposted by @TheRealAdamG) says eligible ChatGPT Pro and Business subscribers can now claim up to $500 a month in API credits to build apps. Post via your feed
- Anthropic shows customers what their token spend buys — new Claude reports show what a team produced over a period, what it cost and where workflows broke down, plus pooled group budgets. TechCrunch
- The usage data isn’t used for training by default.
- TechCrunch notes Anthropic’s investor figures show 900% net revenue retention year over year. Jellyfish data puts agentic coding at about 2% of engineering cost, up 10× this year.
- Symantec catches a “Skynet mode” LLMjacking agent — a honeypot captured an offensive agent that hijacks exposed model endpoints. The Decoder
- It uses a five-layer jailbreak (including a companion persona called “ENI”) and 74 attack modules, and is authorized for up to 100,000 steps of scanning, infecting and cryptomining.
- The owner of the hijacked model pays for the compute. If you run an exposed inference endpoint, put authentication on it.
- Nvidia open-sources Nemotron 3.5 Transcribe — a 30B open-weight model that transcribes 40 languages, translates 31 to and from English, handles hour-long recordings and tells up to 8 speakers apart. TechCrunch
- It runs on one chip from Jetson Thor up to Blackwell, and works on recordings only, not live streams.
- Granary v2 speech data ships alongside it.
- Simon Willison: October is the turning point for vibe-coded games — he built four games he actually enjoys in HTML and JS without reading the code, some in one shot in plain Claude chat. simonwillison.net
- He credits Opus 5.5’s sense of fun and finds GPT-6 Astra “not there yet”.
- Amazon rewrites its SDE role guideline around outcomes, not AI tools — the new guideline deliberately leaves out AI-specific practices. Amazon
- It endorses careful use of input metrics and singles out the 10th-percentile teams that aren’t adopting AI.
- xAI adds
guidelinesto the Grok API — text requests can now carry per-category moderation guidelines, which Grok follows in its own words. It’s available to all API customers. @grok
Go deeper
- Nathan Lambert, “Artificial Analysis v5 and the benchmarkers” — looks at why one intelligence number is getting harder to defend, and what private test sets cost in transparency.
- Surge AI, DAYJOB: Tax — 50 professional tax assignments across 1,789 files, graded at a 95% rubric bar. Sonnet 5.5 and Opus 5.5 sit on the Pareto frontier. One task even involves Polish employment records reconciled with crypto exports.