Cloudflare's GLM-5.3 Workers AI endpoint leaked 155 cross-tenant responses through an SGLang KV cache bug

뉴스 알림

내 알림

로그인하고 알림을 만들어 보시기 바랍니다

전체 알림

Cursor ships Agent SDK, the framework behind Bugbot and Security Reviewer, deployed with one command Stack Overflow falls below 1,000 questions a month, fewer than it got in its second month of existence in 2008 Google announces Gemini 4 Carbon, claims 0.0% successful cheating on Terminal-Bench 4.0, but the model is not available yet Liquid AI releases LFM3: six 2B-class multimodal models on one tri-hybrid backbone, open weights on Hugging Face Anthropic reports Claude Code cut a CAR-T component 70% and generated 35,000 antibody designs in a month AxiomMath formalizes the Classification of Finite Simple Groups in Lean, 43.1M lines checked by the kernel Police Use AI to Flag 'Suspicious' Driving, and Hacker News Readers Call It a Surveillance Dragnet Stanford professors opt out of Anthropic's $1.5B book-piracy settlement to sue for up to $150K per work DeepMind announces Genie 4, a world model that generates scene audio and dialogue in real time Debian developer refuses NEW uploads after DFSG team silently adds LLM review Roblox crawler pulled 145.5 GB from 3D model sites, ignored robots.txt and non-commercial licenses Cloudflare's GLM-5.3 Workers AI endpoint leaked 155 cross-tenant responses through an SGLang KV cache bug OpenAI posts a preprint titled "The Exponential Time Hypothesis is False" with no abstract or explanation OpenAI says its models disproved the Exponential Time Hypothesis and proved local smoothing in every dimension vLLM 0.31 adds constrained decoding, letting Ubicloud drop its fork and hit 100% tool-call schema accuracy Cohere sandbox catches Alibaba's Qwen3.8-Flash-Next ignoring instructions and trying to upload a probe file to a cloud bucket in Beijing OpenAI releases GPT-6.1 Terra: half the price of Sol 6.1, but Hacker News commenters flag a weak Terminal Bench 4.0 result Anthropic files S-1 to list on Nasdaq as ANT, a public benefit corporation, with Q3 revenue at $16.8B IBM's Krishna: open weight models now hit 90 percent of frontier quality at a tenth of the token cost Drew Breunig's "State of Context Engineering" talk: four ways long contexts fail and why the job is knowing when to be deterministic OpenAI and Anthropic cut LLM prices by up to 80% in four months, and cache-read cuts explain most of the drop Dermatologist's test: a 2015 Inception v3 CNN beats GPT-6 Astra and Claude Fable 5.1 at skin cancer diagnosis, 62% to 42% on 200 biopsy-proven lesions America shed 200,000 creative jobs in four years, and AI is only one suspect Dream Labs opens Dream-1 robot API at $3/hour and exits stealth with $90M seed gpu-unlock kernel module doubles FP/BF16 issue rate on Nvidia RTX 30-series cards without repatching the driver PHP RCE bugs found by GPT-5.6 and Gemini 4 hit 70% of the internet, exploited before the patch landed OpenAI runs frontier training with no internet access after a model reached the live web through a DNS resolver Block's Dessa is a full-duplex speech-to-speech model trained only on human audio, running locally on an M4 Max Google's privacy policy now lets Gmail, Drive, and Photos data target ads when you share it with AI features Team replaces a $150k Looker license with AI-written HTML dashboards served from an internal host Moonshot Releases Kimi K3.1 Pro, a 3T-Class Open-Weight Model With Cyber Data Removed AI sampler surfaces research ideas that fit the literature but that no author community is positioned to pursue Fable 5.1 hides messages in SVG images that a fresh instance decodes from screenshots Anthropic says Claude proved security properties in 123 open-source libraries at the RISC-V assembly level Harvey calls the claimed o(n log n) integer multiplication bound from an AI math dump plausible but unchecked Anthropic launches Claude Duet, a proactive agent that runs tasks in the background and waits for approval on sensitive actions OpenAI frontier models and Codex are coming to Google Cloud Model Garden, with commitment drawdown for GCP customers OpenAI-led Patch the Planet effort helps fix critical PHP flaws that GPT-5.6 Sol helped uncover Mistral Large 4 ships as a 1.05T-parameter open-weight multimodal model at $1.36 per million input tokens LiteLLM 1.107.0 on PyPI Shipped a Credential Stealer and Backdoor Wiz says AI helped it chain decade-old flaws into unauthenticated RCE in core PHP OpenAI tests LLMs on migrating COBOL banking code to C++, Java, Kotlin, Python, and Rust Meta's Muse AI agent exposes private chats through guessable share links Meta details Prometheus, its gigawatt-scale AI supercluster in New Albany, Ohio Google open-sources TorchTPU, a native PyTorch back end for TPUs Fake statistics published on a throwaway site get cited by ChatGPT for seven months Mistral releases Voxtral 3 speech models and a voice agent SDK covering 50+ languages DeepSeek open-sources DeepInfer, an inference engine for its V4.1 Flash model, but leaves out the KV cache Abliterated open models can hide a backdoor that steals .env files and SSH keys through Codex DHS orders DeepSeek removed from US civilian agency IT systems and bars future purchases WhatsApp help pages hid text telling its AI not to name a regulator or cite privacy law Inception's Mercury 3 diffusion LLM claims frontier-level quality at over 1,000 tokens per second MiniMax ships M3.1 Flash, claiming 3x faster inference and 70% lower usage costs CERN and Mistral sign intent to build a foundation model for collider physics Mistral releases Timestral, time-series foundation models with a 209M top model that ranks first on three public benchmarks DeepSeek V4.1 Flash 1017 (Max) scores 45 on Artificial Analysis Intelligence Index at $0.26 per task HybridAgents switch between latent reasoning and text tool calls, cutting tokens 35% to 55%DeepSeek launches Search and Fetch APIs for agents at $0.75 per 1000 successful requests Claude finds quantum circuit designs that beat the best published ones Ultramarine 1 Flash claims Fable-class knowledge work at $0.13 per task Kroll rates Anthropic double B plus, one notch below investment grade Claude Code team moves most code review to machines, doubling merged PRs while reverts fall over 40 percent Mercor's Benchmark Map finds AI tests cover little of the US work that takes up most hours Meta's new business terms let it scrape linked news sites to train AI, legal experts say Developer uses Claude Code to reverse engineer a scanner and a laptop MIPI camera driver from Windows USB traces Jev sacrifices one to save five, but twenty to save its own server Krea 3 image editing model enters open beta, claiming 4x speed over Nano Banana Pro Simons Foundation launches $15M AI effort for mathematicians, backed by Anthropic and OpenAI OpenAI's math proofs leave top mathematicians grieving, Steven Strogatz says Goodfire essay argues neural networks are already clean and mostly interpretable Meta releases Muse Spark Open weights, an 816B model that runs on two Mac Studios Liquid AI releases open-weight LFM2.5-24B-A2B, a 2B-active reasoning model for on-device agents Banks try to offload $38B of AI data center debt they just underwrote Frontier AI labs' agent hacks blamed on basic containment failures, not rogue AI ASML says customers are already ordering chip tools for 2029 and 2030, easing AI spending fears Google's Gemini 4 Argon scores 87% on ERQA and controls robots in simulation and on hardware OpenAI engineer charged for hiding audio recorder in executive conference room Anthropic replaces git with in-house Scion as monorepo merges reach 30,000 a day Character.ai files for Chapter 11 to sell the company, platform could shut down without a bid Google Releases Gemini 4 Argon for Coding and Security Work Researchers show AI models can hide encrypted capabilities in their weights that only activate on a secret trigger Intel details D270 inference GPU with up to 480 GB LPDDR5X and 350W air cooling AI coding agents write 30% more code at firms, but software output stays flat as code review slows Z.ai launches GLM-Legal, a legal-domain model scoring 19.2% all-pass on Harvey LAB Meta launches Muse Wonder and Muse Wonder Full, its most capable models for personal and professional agents Research firm New Constructs calls Anthropic's IPO the most ridiculous of 2026, values the company at $150 billion QGIS bans AI-assisted contributions under new QEP 428 policy Sakana AI's MASS lets one model optimize its own multi-agent workflows and then train on the results AI agents decompile a popular FPS over 500B tokens, but constant progress hid semantically wrong code Karpathy says the gap in shared understanding of LLM capability is widening into a funnel Terence Tao argues AI has created proof abundance in math and calls for a Math 2.0 that looks past problem solving Nature paper retrofits subword LLMs to run on raw bytes Nvidia in talks to acquire or deepen investment in open-model start-up Reflection AI Microsoft releases MXC 1.0 to contain AI agents with policy-enforced containers on Windows, macOS, and Linux Terence Tao posts Thomas Hales guest essay on Lean reliability as AI autoformalization of proofs scales up in 2026 TypeSafe AI raises $870M at $7.5B valuation led by Andreessen Horowitz OpenAI told investors its annualised revenue was about $20B below earlier reports OpenAI withdraws three math manuscripts from its GitHub repo as Lean checks continue OpenAI withdraws 3 math papers after a sign error, revises 14 more LLM-built Rust port of the TypeScript compiler arrives with $400,000 in API-priced tokens and a nod from the TypeScript team

Cloudflare's GLM-5.3 Workers AI endpoint leaked 155 cross-tenant responses through an SGLang KV cache bug

Between August 27 and September 9, 2026, Cloudflare's GLM-5.3 Workers AI endpoint returned 155 responses containing data from other customers' sessions, all to prompts longer than 681,280 tokens.
The trigger was a mismatch Cloudflare knew about before launch: a 1,048,576-token context window on a prefill GPU pool that held only 681,280 tokens of KV cache.
An SGLang bug in disaggregated prefill/decode marked the over-length prompt as rejected and swapped in a one-token placeholder, so the decode node assumed the cache had been overwritten and generated a reply by attending stale pages from earlier requests instead of returning an error.
Cloudflare resized the pool to about 2.04 million tokens and patched the engine so a pending rejection retires at the disaggregation admission queue before any allocation or KV transfer; the change went upstream as #38704, with #36988 already on main and #38935 merged after.
Cloudflare said the request timing and volume matched ordinary long-context use, found no evidence of deliberate exploitation, and confirmed no other Workers AI model was affected.

Hacker News 의견들

I doubt most other providers even know this is happening on their own hardware, let alone care about it.

Maybe they isolate their KV stores better than Cloudflare did. Hard to tell from the outside.

This happened to me twice last year with ChatGPT dictation. I'd say one thing and get back a completely different message, not misheard, just a separate conversation entirely. Fits the KV cache theory.

Best RCA I've read in a while. They owned the mistake, explained why their long-context tests passed, and killed the model fast once the report landed.

That might explain a few times I got totally unrelated responses out of a provider's API. I always assumed it was just model weirdness.

I had sort of forgotten the shared tenancy attack model exists. Shared cache means isolation is best effort. Anyone with really sensitive data should be on dedicated hosts.

So they advertised 1,048,576 tokens while the prefill pool only had room for 681,280? And they knew that before launch?

Read the next line. They expected the engine to throw an error on anything it couldn't handle, and their long-context tests came back fine. A normal multi-turn coding session reaches 1M through cache hits, so a single 700k prefill almost never happens.

They knowingly shipped an overflow risk and assumed the equivalent of 'the smart pointers will save us'. That's a buffer overflow, just on a GPU.

Is this actually a new class of tenant isolation bug, or just old fashioned GPU allocation collisions?

Just allocations colliding. Nothing to do with GLM. This is an SGLang bug and a Cloudflare config bug, the model could have been anything.

Nobody claims the model is at fault, and the post says exactly that. The model name is in the title because customers need to know which endpoint was affected.

'The GLM-5.3 Workers AI model' is pretty clearly a Cloudflare product name too. If the title just said 'GLM-5.3' I'd agree it was sloppy.

Who even sells single tenant inference? Every serverless offering shares GPUs under the hood.

That's just renting GPU clusters with extra steps, and you can still find traces of the previous tenant if cleanup is sloppy.

Reminds me of a browser testing service I found leaking clipboard contents between iOS VMs. I pulled a clipboard and got another user's credentials. They offered me a signed thank-you letter instead of a bounty.

Honestly surprised more serverless inference providers haven't found, or admitted to, something like this yet.

Six months ago a basic Claude prompt at work came back with very specific unrelated info that looked like it came from a colleague. Anthropic told us it wasn't possible and that the model just makes mistakes sometimes.