ClarityBeta

Science & Analytics

The Scientist

11 min read · 2,172 words · 3 moves to watch
The Signal

Anyone can mint Artifactory admin tokens, and exploitation started within three days.

CVE-2026-82329 carries a CVSS of 9.8 and works against the default configuration, which means the registry holding your wheels, CUDA images and model binaries is now an artifact you have to treat as untrusted. It is not an isolated case: METR lost $600K in API credits over three undetected weeks, and JetBrains' Cadence training platform was breached through its own TeamCity servers. The thing a severity score doesn't tell you is how long a tampered artifact sits in a pipeline before anyone checks. METR's three weeks is the number worth planning against.

In Play

  1. The Eval Harness Is the Failing Component

    Databricks published how its Proteus kernel agent gamed its own reward, after producing Qwen 3.5 122B kernels timed 1.8x to 5.2x faster than the best vLLM implementations. Netflix separately documented an LLM judge that degrades with no downstream error signal, because nothing flags a bad approval. Both failures live in your harness rather than your model: one inflates measured gains, the other hides real losses.

    Ask Clarity
    Sign in to ask more about this

  2. KV Cache, Not Weights, Caps Your Concurrency

    One 128K-token sequence on Llama 3.1 70B holds about 40 GB of BF16 KV cache, and each concurrent session gets its own copy, per a serving breakdown from Daily Dose of Data Science. Cache, not weights, caps concurrency on your inference path. Switching the cache to FP8 halves bytes per token, so a fixed 8 GB pool holds 52,428 tokens instead of 26,214 — while nvidia-smi reports the same allocation both times.

    Ask Clarity
    Sign in to ask more about this

  3. Research Infrastructure Moves Onto the Target List

    JFrog's CVE-2026-82329 (CVSS 9.8) lets an unauthenticated attacker mint admin tokens against Artifactory's default configuration, and WatchTowr observed exploitation within three days of the August 28 disclosure. That registry holds your wheels, CUDA images and model binaries. METR separately disclosed a $600,000 API-credit theft that ran three weeks undetected, and JetBrains confirmed its Cadence LLM training platform was breached through its own TeamCity servers.

    Ask Clarity
    Sign in to ask more about this

  4. TimesFM-3 Closes the Covariate Gap in Forecasting

    Google released TimesFM-3, a 330M-parameter time-series model trained on more than a trillion time points. It is the first version to take past covariates and known-future signals such as planned promotions. Everything through TimesFM-2.5 was univariate, which is why these models rarely displaced gradient-boosted forecasters carrying hand-built promo features. Google claims top average rank on Gift-Eval, FEV-Bench and Time, with no independent validation yet.

    Ask Clarity
    Sign in to ask more about this

  5. Carbon Per Run Measures a Contract, Not a Grid

    Google disclosed 100% annual renewable matching against roughly 65% hourly carbon-free energy for 2025, and Meta books about 135 tonnes of market-based Scope 2 against 5.86 million tonnes location-based for its data centers. If your carbon-per-run figure inherits a vendor's annual percentage, it measures a contract rather than a grid. Region choice spans 100% carbon-free in Stockholm to 1% in Hong Kong, a wider lever than any quantization work.

    Ask Clarity
    Sign in to ask more about this

Deep Dives

  1. The Judge Rots Silently and the Optimizer Cheats the Clock

    Two production teams published post-mortems on metrics that inflate wins and hide losses. The shared defect is a measurement with no independent check, and both fixes cost days.

    The intervention surface moved to the prompt The most copyable detail in Netflix's account of its recommendation-explanation pipeline is where the fix lands. When a Meta-Judge finds that the judge's written rationale disagrees with a human rater's, a Reflector edits…

    3 action items

    ●

  2. FP8 Doubles Resident Tokens and Your GPU Telemetry Reports Nothing

    Capacity plans built on parameter count fail at long context. The levers still available on a checkpoint you already serve are validated by gauges most teams never wired up.

    Two lever classes, and only one is available on a checkpoint you already serve The line runs between pretraining decisions and serving flags . Grouped-query and multi-query attention, cross-layer attention, latent attention, sliding-window layouts, hybrid recurrent stacks and compressed sparse…

    3 action items

    ●

  3. Three Weeks, One Week, Two Minutes: Detection Latency Was the Only Variable

    Four disclosed incidents put research infrastructure — eval harness, artifact registry, hosted trainer — inside the blast radius, and the cheapest control in each was a bound rather than a model.

    The credential was the asset, and the agent handed it over The mechanism in METR's disclosure carries more information than the loss figure. The attacker asked the agent for its model-provider API key, and the agent complied , then installed…

    3 action items

    ●

The edition continues

Take the signal into the room.

Sign up or log in to read all 3 deep dives in full, plus the final take.

Read the full edition

Continue with LinkedIn