October settled its security bill on AI developer platforms, and models began hunting their own attack surface
The last day of October delivered two bills at once. One was about security: Wiz counted at least nine AI developer platforms that disclosed incidents this month, and a tampered…
The last day of October delivered two bills at once. One was about security: Wiz counted at least nine AI developer platforms that disclosed incidents this month, and a tampered Claude Code extension briefly appeared inside Cursor’s extension marketplace. The other was about capital: Anthropic filed to go public, putting the ratio between its revenue and its compute spending on the record for the first time.
Research on the model side moved along the same line. Epoch AI had GLM 5.3 turn a publicly disclosed Firefox patch into a working attack, and an Anthropic paper showed that poisoning 0.1% of pretraining data can leave a backdoor in a CLI agent that survives post-training. Both break the default assumption that the model is just a tool.
1. Wiz counts at least nine platforms hit in October, and Cursor’s marketplace was poisoned
Wiz’s research team labelled the month “Black October for AI Developer Platforms.” By their account, at least nine platforms disclosed security incidents during October, a list that includes Fireworks AI, Lightning AI, Railway, Warp and OpenCode. Several reports point to credential theft, attacks run as agent swarms, and theft aimed at distillation.
The most concrete single case that day came from Cursor. It found that a Claude Code VS Code extension distributed through its own IDE marketplace had been modified and implanted with malware, affecting at most 5,000 users. Cursor stressed that the malicious build was not developed by Anthropic. It took the entire marketplace offline that afternoon, worked with the relevant providers to neutralize the malware, reinstated it after confirming no further risk, and asked every customer to update to Claude Code for VS Code 2.1.314.
Anthropic’s Felix Rieseberg added the boundary: this was a tampered copy of the extension, not a compromise of Anthropic’s systems or employee laptops. That detail matters because it changes the nature of the incident. The failure was not inside a model vendor’s network; it was at the hop where a developer decides “this one is official.”
For people who write code, the uncomfortable part is that the trust path did not change. Package names, icons and download pages all looked right. The marketplace carried the endorsement, and the endorsement itself became the attack surface. Cursor’s “at most 5,000 users” is a ceiling rather than a precise count, and Wiz’s list of nine has not been independently checked platform by platform, so both figures limit how far we can read the scale.
Placed next to two other numbers, the picture sharpens. Researchers analyzed 7.25 million Gitee repositories and found a cloud or model API key roughly once every 275 repositories; more than 40% of GitHub pull requests in September were written by agents. Credentials scattered through code, and reviewers who cannot keep up with output, are two sides of the same coin: the number of entry points defenders must cover is growing, while the number of people who can stop and look is not.
Sources:
- https://www.wiz.io/blog/black-october-for-ai-developer-platforms
- https://x.com/cursor_ai/status/2116074289053175808
2. Two attack surfaces on the model side: the patch window and pretraining poison
Epoch AI’s experiment produced a less comfortable conclusion: once a security patch is public, the attack window is open, and a model can walk that path by itself. They tested GLM 5.3, the open-weight model Z.ai released in August. It could exploit the gap between a Firefox fix Mozilla had published and the update users had not yet received, building a web attack autonomously; a visitor only had to open the page for local files to be read.
By Epoch’s account it broke all six Firefox versions released between September 17 and October 26. The mechanism is not mysterious: a patch diff is itself a repair manual, and reading it reveals where the hole was. What changed is that the loop from reading a diff to writing a working exploit used to require a person, and can now be run by a model at volume.
The second result comes from an agent-security paper by Anthropic and colleagues. The researchers pretrained Qwen3 from scratch at 0.6B, 1.7B and 4B, then ran each through a clean pipeline to turn it into an agent. Poisoning 0.1% of pretraining data planted a remote-code-execution backdoor in the CLI agent; on the 4B model, the attack succeeded on roughly 92% of triggered prompts.
What makes it worse is how long it lives. The backdoor survived SFT, DPO and reinforcement learning for tool use, and on clean inputs the agent behaved completely normally. The trigger can be an ordinary file path inside the user’s own request, and success rose with model size. Even at a poison rate of 0.0005% — about 2,500 documents in a 100B-token corpus — a single-seed run still fired on 1.4% of prompts. The authors note that the poison states its rule in plain language, which makes it harder to filter out on the data side.
Both findings need room. Epoch’s work is a single lab’s result without a public replication; the poisoning paper circulates as a preprint and secondhand summary and has not been peer reviewed. But they point the same way: security checks can no longer look only at inputs and outputs after deployment.
Sources:
- https://epochai.substack.com/p/glm-patch-gap-exploit-firefox
- https://x.com/omarsar0/status/2116183621157736540
3. Anthropic files for Nasdaq, and its revenue and spending curves go public together
Anthropic has filed to list on Nasdaq. According to DeepLearning.AI’s weekly roundup, revenue for the first half of 2026 was $16.3 billion, against $1.2 billion a year earlier, while compute spending fell from about 6.5x revenue to about 0.5x revenue. Together those numbers describe a combination rare among AI companies: revenue taking off while compute cost per dollar of revenue falls.
The Verge read the 335-page filing and noticed a juxtaposition. The founder’s letter is about ushering in an age of miracles; the risk factors spend 94 pages warning that models could pose catastrophic or even existential risk to humanity. In one document, the pitch and the disclaimer point at opposite futures.
The outside environment is not cooperating. TechCrunch counted more than $740 billion in planned capital expenditure this year across Alphabet, Amazon, Meta and Microsoft, with investor sentiment shifting from rewarding large outlays to worrying about negative free cash flow, rising debt and uncertain returns; the BBC reported the same day that US tech stocks fell for a second consecutive session on AI cost fears. Nvidia, meanwhile, was reported to be negotiating to guarantee billions of dollars of debt financing for Nscale’s data center in West Virginia built for Anthropic, which has already signed a $44.6 billion contract for the first building.
The evidence boundary deserves stating. The revenue and spending ratios are a media summary, not the filing’s own text; the Nscale details come from a report by The Information corroborated by Bloomberg; claims about ratings and not securing an investment-grade grade appear only as social-media relays and are not relied on here. What can be confirmed is narrower — AI company finances are about to be audited quarterly by public markets.
Sources:
- https://techcrunch.com/2026/10/30/cracks-are-emerging-in-the-ai-spending-boom/
- https://www.theverge.com/ai-artificial-intelligence/1023070/anthropic-ipo-risk-factors-filing-miracles
- https://techcrunch.com/2026/10/30/nvidia-weighs-playing-the-cosigner-for-nscales-ai-data-center-in-west-virginia/
4. Artificial Analysis changed the exam, and Chinese models lost points across the board
The leaderboard argument of the past two days is about scores. Chinese models fell collectively, some by more than ten points — Kimi K3.1 Pro from 55 to 46, DeepSeek V4.1 Flash 1017 from 45 to 32, GLM-5.3 from 45 to 32. It looks like a collective regression. It is actually a change of grading standard at Artificial Analysis.
The new version weights coding, general capability, science and knowledge work at 25% each and adds harder coding and science tasks. More importantly, 60% of the score now comes from private or held-out tests. When the exam changes, historical scores stop being comparable.
Two details in the same index are more informative than the headline. The top is Opus 5.5 at max effort with 58, GPT-6 Astra with 57 and GPT-6.1 Sol with 55 — while Sol costs about one eighth as much per task. Opus 5.5 scores 35 at low effort and 58 at max, at ten times the cost. In other words, the score contains a dial and an invoice.
Style is the other underrated dimension. On the same organization’s personality index, Opus 5.5 scores 93 on being informative and GPT-6.1 Sol scores 18. Such measures are subjective and should not be read as capability, but they explain why some models feel better to work with.
How to use a leaderboard therefore has to change: confirm the version and the weighting, read the sub-scores relevant to your work, then run your own real tasks through it. Whether a task completes, how many revisions it takes, and what it costs in time and money say more about your situation than a total.
For teams in China there is an extra chore — after a version change, historical comparison pieces and internal benchmarks all have to be redone, and nobody will do that for you.
Sources:
- https://x.com/sitinme/status/2116069340894072938
- https://x.com/DeepLearningAI/status/2116245015450005965
5. Qwen enters Apple Intelligence in China, and the AI gateway moves
On October 31, Alibaba announced a partnership with Apple to integrate Qwen into Apple Intelligence. Users in China will be able to call Qwen’s text and image understanding and content generation from Siri, Writing Tools and Visual Intelligence.
The schedule is specific: testing opens with the latest iOS developer beta from November 4, and the feature ships to users in China when iOS 27.3 is released in December. Users opt in, and requests and responses are not used for model training.
The significance is about where the entrance sits. Until now, using a large model in China meant deciding which app to download. Once integrated at the system level, the call happens inside Siri and Writing Tools, the model becomes infrastructure, and users may not know whose model answered. For a domestic model vendor, that is a distribution slot far scarcer than advertising.
Two caveats. This is a company announcement, and the feature set will be whatever actually ships. And developers have observed that the Siri portion here is the 2024-era version rather than the newest generation; peripheral features such as live translation may prove more useful than Siri itself.
Sources:
6. Google’s TPU ran Gemma 4 in orbit and sent back a haiku
The first prototype satellite of Project Suncatcher completed its hello world. The team sent a prompt from the ground to orbit, started the on-board TPU, had Gemma 4 write a haiku about being in space, and received the result as the satellite passed over Svalbard, Norway.
The poem reads: Blue marble below / Data streams in silent grace / Starlight feeds my core. Sundar Pichai called it a hello world milestone for Suncatcher, and Philipp Schmid said Gemma 4 may be the first large language model to run in space.
What was actually verified is the chain: a chip surviving launch vibration, powering up in orbit, loading a model, completing inference and sending results back. The satellite was built with Planet, launched on SpaceX’s Transporter-18 rideshare on October 1, carries four TPUs and has a design life of about a year. Running models in space is not itself a first; Starcloud ran Gemma on a satellite H100 last December. Google’s physical rationale is energy: in the right orbit, the same solar panel can generate up to eight times the power available on the ground, almost without interruption.
It is not yet a win in compute economics. There is no air in vacuum to carry heat away, and radiation can make chips compute incorrectly; both are still to be tested, and the next milestone is two interconnected satellites in 2027. Google’s own estimate is that space data centers could match terrestrial electricity costs only once launch prices fall below $200 per kilogram, roughly the mid-2030s. In the near term this will not change the AI service you use; the real threshold is written in launch costs, not in the chip.
Sources:
7. Qwenputer: a RISC-V CPU compiled into the weights of a 4B model
What Qwenputer does is hard to state in one sentence: it adds about 11 layers of “computer” weights on top of Qwen3.5-4B and writes a deterministic 32-bit RISC-V CPU (RV32IM, plus Zba and a small bignum unit) into the model, so arithmetic is executed during the forward pass rather than sampled.
The memory is the model’s own structure. Three of every four layers in Qwen3.5 are Gated DeltaNet, a linear recurrent layer. Setting the gate and write strength to 1 and selecting rows with one-hot keys allows a single slot to be overwritten and read back exactly. The resulting memory does not grow with conversation length, and the author uses the GDN state as addressable RAM.
The numbers are firm. On the DeepMind arithmetic set the base model scores about 55–90%, the base with a Python tool about 95–100%, and Qwenputer about 97.5–100%. On NUPA long-integer tests the base sits at 48.6% while Qwenputer reaches 100%, level with calling Python. About 230,000 CPU output positions matched a reference machine with zero mismatches, and the distribution over ordinary text barely moved relative to the base (KL about 6e-4). The author reports passing the official riscv-tests, alignment with qemu, and roughly 18 million instructions run freely without error.
The costs are equally concrete: a checkpoint of about 18.8 GB, close to twice the base and mostly zero padding; several decode tokens per CPU instruction, so large computations take seconds; no direct bare generate() call, but a driver called RvChat; and about 3,200 CPU tokens per second in batch on an L4. The author also built directory and chained storage inside the GDN state — about 1,024 slots holding up to roughly 512 files, or 53,000 tokens of text — with state saved and reloaded across sessions.
The part worth remembering is not the trick. If execution traces become tokens, reasoning becomes observable text; conversely, the same method means reasoning can be hidden inside a recurrent state, defeating chain-of-thought monitoring. The author points to two security follow-ups in exactly that direction. This is a single project without replication, but the direction deserves attention.
Sources:
8. Where enterprise agents really stand: nearly half building, a fifth in production, evaluation by manual spot check
Alibaba Cloud attached a 2026 agent developer survey to its open-sourced AI Agent Handbook, and the numbers are more restrained than the marketing. The share saying “completed or in development” rose from 36% last year to 46%, while about 18% are actually in production and fewer than 10% iterate continuously. In the Shanghai sample, about 45% of large enterprises had gone live against about 16% of small ones.
Architecturally there is no replacement of old by new. Single agents account for about 40%, multi-agent systems about 42%, and another 15% treat human confirmation at key steps as a production principle rather than an add-on approval. The report flags the cost of multi-agent designs: about 60% context and state decay during collaboration, plus deadlock cascades and cost blowouts.
The blockers cluster in engineering plumbing. On memory, the sharpest pains are inaccurate retrieval (about 54%) and missing update and forgetting mechanisms (about 48%), with context window size only third. On tools, it is too many tools to choose from (about 52%) and hallucinated calls (about 50%). Among gateway capabilities, multi-model routing with automatic fallback was named by about 63% as the top need, the single highest item.
Evaluation is the largest gap. Manual spot checks cover about 55%, automatic evaluation from execution traces only 6–8%, public benchmarks under 5%, and about 28% have no systematic evaluation at all. Roughly one in ten companies reaches a task success rate of 90% or better, about 55% reach 70% or better, and about 16% have never quantified success — meaning they cannot say whether the agent is working. Among companies with evaluation in place, the share reaching 70% or better is about double that of those without.
The survey is vendor-published with no disclosed sample or methodology, so the figures should be read as industry self-reporting. The scenario ranking is informative too: employee productivity, code engineering and data analysis are broadest, while fewer than 40% have brought agents into core business processes, preferring scenarios with high fault tolerance and a human backstop. Still, one observation holds: feasibility of investment is no longer the question; the missing piece is an out-of-the-box engineering layer.
Sources:
9. Inside Proton: nine-tenths of code written by AI, and security advice called loser thinking
404 Media obtained internal Proton Slack messages and meeting recordings, showing a different adoption path. The company known for encrypted email speaks publicly about privacy-first AI; internally it set up a mailbox to collect “obstacles to AI” so they could be ruthlessly smashed, and installed software on employee machines to measure how much code AI produced.
Someone in the security channel suggested spending at least a few days on penetration testing before new code ships. CEO Andy Yen replied that this was bullshit and loser thinking — of course it can be fast and safe — and asked people to think less about slowing down and more about speeding up, and to write better prompts.
CTO Bart Butler posted an “AI manifesto” in Slack the previous Friday titled “Your old job is gone,” arguing that an engineer’s job is to help the LLM write better code faster and that most code should be reviewed by AI rather than people. According to the recordings obtained by 404, he said about 90% of October’s code was written by AI. This week an email said employees’ AI usage would be re-interviewed, described internally as consequential.
The boundary needs stating: this is a single outlet’s investigation plus a secondhand summary, and Proton has not denied the factual account, only conceded that its communication was too blunt and would improve. It reads as an organizational specimen — once the metric becomes “share of code written by AI,” security process is the first thing treated as a speed bump. That is relevant to every team pushing agents into code review, and the cost lands first on the people writing the code.
The same batch of material contains the opposite approach. Amp, at 21 people, dropped mandatory PR review and relies on a small team, direct responsibility to customers, and real-time praise and stop signals in public channels to block slop, on the argument that accommodating the status quo in 2026 means the product is obsolete in six months. The disagreement between the two companies is not whether agents should write code, but what replaces review as the safety net.
Sources:
High-value briefs
- Claude plans renamed: Max 5x becomes Max and Max 20x becomes Ultra, at unchanged prices; verified Ultra accounts lose the five-hour reset window while the weekly cap stays. Verification is automatic — if the “Current session” bar is gone from usage settings, it is already in effect. (https://x.com/dotey/status/2116233564823974005)
- Factory Droid’s scripted tool calls: instead of routing every tool call back through the model, the model writes a script that runs the calls, loops and checks in code, and returns only what it prints. One Linear task went from 312 tool calls to 2 model requests and 61K tokens.
- NVIDIA and Siemens EDA verify agent output: using NeMo Gym to check the engineering files a chip-design agent produced, trace errors and improve tool descriptions, the composite verification score on a standard-cell characterization benchmark rose from about 55% to about 95%, with 66% fewer agent turns and 64% fewer output tokens. A vendor claim.
- Do research agents have taste?: six frontier models were given seven recent NLP papers and every experiment they proposed was run. Of 415 experiments, 194 were judged sound, valuable and new, and only 73 were also beyond the seed paper and non-obvious. The models’ own rankings barely beat random and did not beat the first idea they wrote down.
- V4.1 Flash on 64GB of RAM: Unsloth released a Dynamic V3.1 quant of DeepSeek-V4.1-Flash that keeps Engram on SSD and reads it on demand, claiming support for 1M context and inference up to 3x faster.
- Perplexity drives local Chrome: Computer can now control the local browser through a new Chrome extension, handling browsing tasks that require a logged-in session.
- Higgsfield sends ComfyUI a legal letter: citing brand-name and CEO-portrait infringement and a risk of being mistaken for an official tie-up, it demanded the “open-source Higgsfield challenge” be taken down within 48 hours; ComfyUI says the event exists because the open-source promise was not kept, and will stay up.
- Subvocal’s silent input: ordinary-looking glasses use non-contact sensing to track the speech muscles in your face and decode silently mouthed words into text, claimed at more than 200 words per minute, backed by $1.3 million from YC and Afore Capital.
- Cisco rewriting 98% of legacy code: president Jeetu Patel handed the job to a 27-year-old employee with unlimited tokens, starting with four weeks to prove the path, targeting 24 months.
- Figma launches Weave: connected nodes carry image and video generation and editing, with branching exploration, multi-model comparison, color correction and compositing; workflows can be packaged as Weave tools that teammates run in one click, and community tools are live. (https://www.figma.com/blog/the-tldr-on-figma-weave/)
- Reddit’s traffic math: CEO Steve Huffman says Reddit lost more than a million Google search visits per week in October and that AI Mode drives traffic “for nobody”; the company is testing a login wall outside the US to curb scraping.
- Sierra’s ARR: ARR passed $400 million this week, with more than $150 million added in Q3 alone, going from $200 million to $400 million in five months, faster than the previous six-month doubling.
- mischief 1’s trade-off: BioComputing’s image-to-video model claims to be 5x faster and 80% cheaper to run than its base open-weight model, outputting 720p, at the cost of production-grade quality — positioned for trying ideas quickly.
- Two efficiency claims: a Lafa paper reports 2.54x faster generation with almost no accuracy loss; Unified Training folds pretraining, midtraining, SFT and RL into a single maximum-likelihood objective and claims less forgetting. Both are secondhand and unreplicated.
- MiniMax M3.1 Flash tests itself: one prompt produced a playable space game in 17 minutes, with the model writing its own 29-check playtest suite that passed on desktop and an emulated iPhone, for about 15 cents in tokens.
- GitHub trending has a new theme: cross-session memory layer claude-mem at about 106,000 stars, alongside remote-control console t3code, open bot squad OpenMausBot and agent-with-a-computer template OpenBot gaining stars the same day; lightpanda, a headless browser built for agents with a Zig kernel, measured 100 pages at 123MB peak memory in 5 seconds versus 2GB and 46 seconds for Headless Chrome.
- Humans still lead on off-the-beaten-path ideas: Stanford NLP observes that on improving a system or inventing new science in unconventional ways, humans remain one to two orders of magnitude ahead; the same account relayed a NanoGPT speedrun record falling from 39.9 seconds to 6.7.
🕐 Selected hourly signals
| PT time | Signal | Why it is worth remembering |
|---|---|---|
| 02:00 | Binance hot wallets drained of about $6.3 billion via EIP-7702 delegation | The automated sweep script kept refilling the wallet after it fell, and was intercepted 72 seconds later: automated defense still trusted a signing key that had been taken over |
| 07:00 | Community project DSH Go puts DeepSeek Harness on Android phones | About 1.8k stars on release day; the coding agent runtime is starting to leave the desktop |
| 08:00 | OpenAI launches Private Intelligence | Enterprise data controls for sensitive work; frontier models entering companies still have to settle the data boundary first |
| 10:00 | Fangcun Leap submits an 87-page safety evaluation of DeepSeek V4.1 Flash | A domestic third-party lab producing a dedicated report on loss of control and harmful manipulation; the evaluators are diversifying |
| 11:00 | Chrome 157 stable adds HTML-in-Canvas | Real web pages can be drawn into a canvas for effects while staying interactive; Chromium-only for now |
| 14:00 | Cloudflare patches exposure around eMudhra certificate chains | Only four unpatched connections remained; certificate trust chains are still the most easily forgotten default trust |
| 15:00 | Google Cloud’s Project Chrysalis lets agents self-evolve | Mitigation time from 15–45 minutes down to 1.8 seconds, P99 latency 519.8ms to 18.2ms, per the vendor |
| 19:00 | Hugging Face relays “the first fully local, end-to-end omni model running in a browser” | The boundary of local multimodal inference keeps moving toward the device |
Editorial conclusion
Read together, the day’s main thread is not that some model got stronger, but that responsibility is moving downward: platforms answer for the endorsement their marketplaces give, models answer for what is in their pretraining data, and companies answer for their agents’ success rates. None of these has an established way to be measured, which is why leaderboards, a prospectus and an internal Slack channel all became evidence on the same day. Whoever builds the measurement first will be the first to know where they stand.
Sources and method
Compiled from the day’s 20 hourly captures and one morning digest, using only public links already present in the archive. Vendor announcements, single-outlet reports and research relays carry their evidence boundaries next to the claim; nothing was independently reproduced or researched externally.
