Atlas News
RSS

53

2026-11-22Daily

9 stories selected7 source clusters

New Riemann proof drafts and fresh scrutiny of prompt leaks

Two proposed Riemann proofs enter review, while MedChemBench tests medicinal-chemistry reasoning. Qwen4 Max joins Together AI, and AWS Quick expands conversational flows and coordinated agents.

Together’s incident report and Dragos’s tracking experiment examine how prompts travel through service chains. Prompt design, shared caches and output costs round out the engineering updates.

01

Research and science

2 stories

  1. 2026-11-20Levent Alpöge / Ravi Vakil

    Alpöge claims a Claude-assisted Riemann proof as a competing draft faces questions

    Anthropic mathematician Levent Alpöge says he and Claude have found a proof of the Riemann Hypothesis and the Generalized Riemann Hypothesis, releasing formalized material in 14 data packages. He says the formalization is complete but plans to understand, simplify and write up the argument for human readers. This remains the author's proof claim.

    Stanford mathematician Ravi Vakil released an independent draft by Alessandro Masullo the same day, describing it as entirely human work that Brian Conrad had reviewed for months. About eight hours later, Vakil reported that people using different AI models had identified problems. He said those issues must be resolved before community acceptance and made no claim about the strategy's correctness.

  2. 2026-11-20OpenAI

    OpenAI introduces MedChemBench with 99 medicinal-chemistry problems

    OpenAI introduced MedChemBench, comprising 99 problems and 761 distinct molecular structures drawn from 117 papers and patents. It grades structure reconstruction, evidence-based explanations, compound ranking and property prediction. At xhigh reasoning effort, GPT-6.2 Astra earned 84.5% of the available credit, compared with 83.3% for GPT-6 Astra and 82.0% for GPT-6.1 Sol.

    Tasks connect molecular structures and experimental findings to tradeoffs such as activity, permeability and metabolism. Each response is a single attempt without tools, feedback or new experiments, and prior exposure to published sources cannot be ruled out. The scores measure reasoning over historical evidence; they do not establish an ability to discover drugs or improve laboratory decisions.

02

Products and workflows

3 stories

  1. 2026-11-21Together AI

    Together AI launches Qwen4 Max and demonstrates a whole-repository bug fix

    Together AI added Qwen4 Max to serverless inference with a one-million-token context window. In its demonstration, one request contained all 344 Tornado files, 652,519 input tokens and a bug report, without tools or retrieval. The model returned a patch after 72 seconds; it passed two targeted regression tests and the full 1,673-test suite.

    This vendor demonstration fixed one WebSocket error-logging issue when a connection closes, rather than every bug in the repository. A 30% discount runs through November 22: input, cached input and output cost $1.05, $0.0525 and $3.15 per million tokens respectively. The demonstrated request cost about $0.69. Requests must stream, and thinking is enabled by default with its tokens billed as output.

  2. 2026-11-20AWS

    AWS Quick adds conversational flows and coordinated specialist agents

    AWS Quick now lets users describe, run and refine business flows in chat while viewing a diagram of the steps. Flows combine agents, skills, tools and data with deterministic calculations and checks, supporting longer tasks and larger data volumes. Several specialist agents can work on one request concurrently, with their findings combined into a single response.

    The refreshed web app brings actionable email, messages and a daily briefing to the home screen, and long-running work continues after the tab closes. New Meta ads and Amplitude connectors respect existing data permissions and ask before write actions. Skills can also be managed on the web; prompt guardrails remain a preview for Professional and Enterprise plans.

  3. 2026-11-21OpenAI

    OpenAI's GPT-6.2 guide explains prompting changes and history-edit behavior

    OpenAI's GPT-6.2 guide recommends reviewing system prompts during migration. Astra is more concise by default, so additional instructions to be brief may be counterproductive. For long tasks, the guide calls for a clear intended outcome, reasonable assumptions the model may make and decisions reserved for the user. Its coding guidance emphasizes reusing existing logic and adding tests when needed.

    The guide also says the Responses API drops earlier reasoning after edits to conversation history, certain model switches, or changes to the end user or organization. This can cause repeated reasoning and higher token use. Reusable instructions can go in a developer message at the start of input to keep the prefix stable; top-level instructions suit guidance for the current request.

03

Data and security

2 stories

  1. 2026-11-21Together AI

    Together AI discloses prompt-sample theft and defaults new accounts to zero retention

    Together AI says an unauthorized party downloaded prompt samples from a private Hugging Face repository. The samples involved MiniMax M3, GLM-5.2 and Kimi K2.6 requests, mainly from June 24 and July 4, 8 and 14. An AI coding agent had committed an access credential to the public xorl repository; it was obtained during exposure despite subsequent detection and rollback.

    The party then searched the samples for other credentials and tested them, including five canaries from Dragos's experiment. Together is working with external forensic investigators and law enforcement and is still reviewing affected content. It says it has found no evidence of additional prompt leakage beyond this incident. Changes include scoped or short-lived credentials, scanning every pull request, and zero data retention by default for new accounts.

  2. 2026-11-19Dragos

    Dragos traces AI-service leaks with nearly two million canary-bearing prompts

    Dragos researcher HD Moore reports sending nearly two million prompts with unique links or test credentials to AI services between July and November 2026, then monitoring subsequent access. The study found signs of compromises at at least three providers, prompt retention by inference resellers, and failures to honor settings such as zero data retention.

    The credentials led to researcher-controlled systems and empty AWS accounts, so the prompt count is not a count of real victims. In the largest incident, a few hundred bait keys were used, generating tens of thousands of AWS API calls. Comparing requests to the same model through different resellers helped distinguish model-related behavior from leakage tied to a particular service route.

04

Engineering and practice

2 stories

  1. 2026-11-20AWS

    AWS adds a distributed KV-cache skill that checks savings before deployment

    AWS added a distributed KV-caching skill to its Agent Toolkit for coding agents such as Kiro, Claude Code and Codex. It evaluates and configures Amazon ElastiCache, starting with read-only inspection of GPU resources and assessing models, prompt length, repetition and routing. Shared cached attention states can prevent servers from recomputing the same context.

    KVTrace records vLLM or SGLang cache events and compares capacities offline without storing prompt text or token IDs. Block hashes still derive from content, so sensitive traces require the documented anonymization option. The workflow stops when projected savings do not cover cache costs and checks output consistency before production. Freed GPU time becomes monetary savings only through fewer servers or more traffic on the same fleet.

  2. 2026-11-21Sean Goedecke

    Sean Goedecke argues visible-answer caching targets a small share of agent output

    Engineer Sean Goedecke asks whether models could emit code-line or conversation references for an application to expand, reducing repeated text generation. In his own agent histories, visible answers account for about 6% of output tokens, reasoning for 50% and tool calls for 44%. Even eliminating two-thirds of visible output would therefore reduce total output by only about 4%.

    He argues that reducing work in reasoning and tool calls is more likely to affect costs than compressing final answers alone. These figures describe personal usage, not a cross-model benchmark. He also cannot inspect complete frontier-model reasoning and remains uncertain about repetition within it. The analysis separates shorter-looking answers from lower token use across a whole task.

Updated Issue date: 2026-11-22

Subscribe

One brief at a time, only when there is something worth your attention. Unsubscribe anytime.