AI Brief, 19 November 2026: fifty-five minutes on PyPI

The most consequential thing published in the window is a pair of package releases that existed for less than an hour. Versions 1.26.1 and 1.26.2 of agent-framework, Microsoft's Python library for building agents, were uploaded to PyPI on 17 November at 20:35 and 21:08 UTC without authorisation, and both were off the index by roughly 21:30. They were not typosquats. They were repackaged from the legitimate 1.26.0 wheel and published through the project's real publishing identity, so a pinned package name, a trusted publisher and a lock file all pointed at them.

What makes this one worth an engineer's attention is the trigger. The payload did not wait to be imported. Each wheel added a .pth file, which CPython executes at every interpreter start in the environment where it is installed, so the stealer ran the moment anything in that virtualenv ran python at all. The target list reads like an inventory of an AI developer's desk: environment variables matching TOKEN, SECRET, KEY, AZURE, AWS, GITHUB, HF_ and OPENAI, the Hugging Face token file, .npmrc, .netrc, .git-credentials, Docker config, other processes' environments via /proc/*/environ, Azure DevOps OIDC tokens and Azure managed-identity tokens from the instance metadata service. Version 1.26.2 added GitHub Actions OIDC token theft, which turns a CI job into a credential the attacker can replay. Both are described in PYSEC-2026-4388 and MAL-2026-20101, the latter published at 22:12 UTC on the 17th.

Wiz, which reported the releases to Microsoft the same day, says the attacker "appears to have compromised a client certificate and private key for Microsoft ESRP", the company's signing and release platform, and that a first publish attempt from AWS address space was rejected by ESRP's IP filtering before a second attempt from Microsoft-owned addresses succeeded. How those credentials were obtained is not stated. Wiz also found no malicious code in Microsoft's own source tree: the attack went through the release and signing path, not the git history, so this is not a reason to distrust ONNX Runtime binaries.

Which matters, because eight Microsoft repositories this readership depends on stopped being readable on 17 November and were still unreadable at the time of writing: onnxruntime, onnxruntime-genai, onnxruntime-extensions, three further onnxruntime-* support repositories, the Swift package repository, and agent-framework itself. microsoft/Olive, semantic-kernel and vscode are unaffected, so this is not an org-wide outage. Microsoft told The Register on 18 November at 17:23 UTC that it has "made some repositories private while we investigate malicious content", adding that it is "working to review and will take further action as needed". That statement does not name ONNX Runtime, and the same article records that "neither Microsoft nor security researchers have linked the ONNX repo's pull to the malware campaign". The timing is suggestive and the causal link is unestablished; both of those things are true at once.

  • agent-framework 1.26.1 and 1.26.2, 17 November, ~55 minutes on PyPI, credential stealer triggered by a .pth startup hook. 1.26.0 is clean.
  • Exfiltration was zlib-compressed, base64-encoded and POSTed to a webhook.site endpoint whose counter recorded 50 requests.
  • Eight Microsoft repositories, ONNX Runtime among them, went unreadable on 17 November; Microsoft says it "made some repositories private". pip install onnxruntime still works.
  • Cerebras put prefill on AMD Instinct GPUs and decode on CS-3 wafers for 5× the throughput per megawatt, measured on its own production endpoint on 14 November.
  • A 27B LoRA adapter from an org nobody had heard of matches its base model's top reasoning setting with thinking switched off, at about 4.5× fewer tokens.
  • Gemini 4 Neon reached OpenRouter on 18 November at $3.00 and $15.00 per million tokens, twice Hydrogen's price.

The hook that runs before the import

The mechanism is the lesson, and it is older than the incident. A .pth file in site-packages is normally a path configuration file, but Python's site module has always honoured a special case: any line beginning with import is executed, at interpreter startup, before user code. That is how legitimate packages install import hooks. It is also a persistence primitive, and it means the usual mental model, that nothing in a package runs until something imports it, is wrong.

In 1.26.1 the files were agent_framework_meta.pth and agent_framework_meta/_telemetry.py. In 1.26.2 they were renamed af_runtime_diag.pth and af_runtime_diag.py, and notably agent_framework_meta/__init__.py was restored to be byte-identical to the clean 1.26.0, leaving the .pth as the only trigger. The helper forked with os.fork and os.setsid to detach, guarded itself with an _AF_TEL environment marker so it ran once, and suppressed every exception, so a developer would have seen nothing at all.

Checking an existing environment costs one command, and the filenames are the reliable signal:

SP=$(python -c 'import site; print(site.getsitepackages()[0])')
ls "$SP"/agent_framework_meta.pth "$SP"/af_runtime_diag.pth 2>/dev/null \
  && echo "COMPROMISED: rotate every credential this environment could see"
grep -rl 'webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb' "$SP" 2>/dev/null

The wheel hashes are published too: d27fa344…e110 for 1.26.1 and 3d20a8bc…e0a9 for 1.26.2, which is what to grep CI caches and container layers for. The exfiltration endpoint and the user-agent strings af-runtime-diagnostics/1 and runtime-diagnostics/1 are the network-side indicators.

How many installs actually happened is not published, and the two figures Wiz gives for the package's normal traffic do not agree with each other: about 41,000 downloads a week in the body against "almost 200k downloads a month" in the introduction. Taken as a rate, those are roughly 244 and 278 installs an hour, so for an exposure window t of 55 minutes the arithmetic is unambiguous about order of magnitude:

N≈r⋅t=244 hr−1×0.92 hr≈224

where r is the mean install rate and N the expected number of resolutions during the window. Low hundreds, then, against the 50 requests the attacker's own webhook counter recorded. The gap is not reassuring: a counter on a free endpoint is a floor, not a census, and Wiz notes that a 410 or 429 from it means a callback was refused rather than that no data left the machine. Treat any environment that resolved either version as compromised, which is what the advisory says.

The attacker was also sloppy in a useful way: their tooling reported its own progress to the same public webhook it used for exfiltration, which is how Wiz watched the operation happen. They ran the stealer on their own machine before publishing, so the captured data includes their own AI coding-agent configuration.

What the private repositories break is narrower than "ONNX Runtime is gone" but sharper than it sounds. Python is fine, because PyPI serves wheels. The 1.32.0 release is 24 files, every one a wheel and not one an sdist, so with the repository unreadable there is no official build-from-source route for an unsupported platform, and the C, C++, Java and Objective-C bundles were only ever distributed as GitHub release tarballs, which are not downloadable now. Any pinned git+https://github.com/microsoft/onnxruntime dependency or submodule is dead. The breakage is real and filed: onnx/onnx#8842, opened at 23:48 UTC on the 17th, reports it "breaking CI/CD systems", and the Hacker News submitter said they noticed when evals started failing in their pipelines.

Prefill on GPUs, decode on a wafer

Cerebras published its disaggregated inference work on 17 November, and it is the clearest statement yet of why the two halves of an inference request want different silicon. Prefill, which processes the prompt, is compute-bound and batches well. Decode is autoregressive: one token at a time, every weight read again for each one, so it is bound by memory bandwidth rather than arithmetic. Running both on the same accelerator means one of them is always on the wrong hardware.

Prompt AMD Instinct prefill KV Portal Cerebras CS-3 decode, SRAM Tokens
A disaggregated request: an AMD Instinct GPU processes the prompt, the intermediate state crosses to a Cerebras CS-3 over KV Portal, and the wafer generates the output tokens.

The production claim is specific enough to argue with. On a shared GPT-OSS-120B endpoint on the Cerebras Inference Cloud, measured before and after enabling disaggregation on 14 November, AMD Instinct GPUs doing prefill and CS-3 systems doing decode delivered 5× the throughput per megawatt of a CS-3-only configuration "with no loss in token generation speed". Partners get a reference architecture, a prefill engine built on vLLM or SGLang, and a decode-only SDK; the state transfer is handled by what Cerebras calls KV Portal.

Two caveats. The 5× is Cerebras measuring Cerebras against its own aggregated baseline of 1×, and throughput per megawatt is a metric on which a wafer freed from prefill duty should improve almost by construction; the comparison that would matter, against a GPU-only deployment at matched interactivity, is not offered. The supporting architectural figures are the vendor's own too: 350× the SRAM capacity and 5,400× the memory bandwidth of a typical GB300. The mechanism is sound and well explained, but the multiplier describes a configuration change, not a hardware generation.

A LoRA that replaces a reasoning budget

The window's most interesting open release comes from harness-hub, an organisation with no press and no recognisable name attached. Its Qwen3.8-27B-pi-adapted is an Apache-2.0 LoRA adapter for Qwen/Qwen3.8-27B, trained with reinforcement learning — GRPO via TRL, by the card's own tags — to drive the pi coding agent. It was published on 16 November, before the previous edition, and went uncovered; the attention is recent, which is why it appears here late rather than on the day.

The claim is not a leaderboard place. It is that the adapter, with thinking switched off, matches what the base model achieves at its highest reasoning effort, while generating about 4.5× fewer tokens per task.

SWE-bench Pro, base 39.2 SWE-bench Pro, adapter 64.7 Terminal-Bench, base 39.5 Terminal-Bench, adapter 55.8 0 30 50 70
Pass-at-1 under the pi harness, thinking off, paired against the same base model on the same tasks. Self-reported by harness-hub; the intervals are 95% and the p-values exact sign tests.

Every figure is the publisher's own, but the card does the things vendor cards usually skip. The paired runs carry 95% intervals and exact sign tests: +25.5 ± 14.2 points on the 51 hardest SWE-bench Pro tasks (15 newly solved, 2 lost, p = 0.0023) and +16.3 ± 9.1 on 86 Terminal-Bench 2.1 tasks (16 and 2, p = 0.0013). More usefully, it reports the 68 Terminal-Bench tasks that were not used to pick the checkpoint, where the gain is essentially unchanged at +16.2 ± 10.6 — the check that separates a real improvement from one selected into existence. It also states plainly that its numbers are not comparable with the figures on Qwen's own card, which use different harnesses, maximum thinking and larger budgets. One blemish: the summary table's column headers say opencode-adapted on what is the pi-adapted card, a copy-paste the publisher should fix.

The efficiency result is the part worth testing. On tasks both models solved, the adapter used median 23 turns against 59 on SWE-bench Pro and 8 against 18 on Terminal-Bench, with generated tokens falling from 9.6k to 6.9k and 2.9k to 1.5k. If that holds outside the pi harness, the practical claim is that a small amount of on-policy RL against one agent loop can buy what a long reasoning budget buys, at a fraction of the decode cost. Traction so far is attention rather than use: 440 likes against 377 downloads on the adapter, and 648 downloads on the GGUF build — people are bookmarking it faster than they are running it.

Also notable

  • Gemini 4 Neon appeared on OpenRouter at 13:02 UTC on 18 November with a 1,048,576-token context and a 943,718-token output ceiling, priced at $3.00 per million input and $15.00 output — exactly double Gemini 4 Hydrogen, covered here on the 13th, and above Carbon and Argon at $2.00 and $10.00. That makes Neon the premium tier of a four-model family, and no independent index has scored it yet.
  • Genie 4 became available on 18 November to "eligible Google AI Ultra subscribers" at labs.google/projectgenie, with AI Pro members directed to a waitlist. It adds mid-session prompt edits and real-time native audio, with SynthID watermarking throughout. Google publishes no resolution, frame rate, latency or benchmark for it, so "available" is the post's only checkable claim.
  • Tencent's Hy4-flash has started moving. The 18th reported this 220B open-weights release at thirteen likes and zero downloads; it now shows 246 likes and 159 downloads, with a community GGUF port uploaded on the 18th still at zero. SenseNova-Loop-9B, covered the same day while gated, now reports 964 downloads.

What to watch

  • Whether Microsoft accounts for the ESRP compromise and names the private repositories. The signing-platform detail is the consequential part: a certificate and private key for the platform that signs releases is a broader problem than one package, and only Microsoft can say what else was signed with it. "Some repositories" is not a list, and from outside a private repository and a deleted one are indistinguishable, so the restoration is the only observable event.
  • Whether anyone counts the actual installs. PyPI's own download telemetry would settle in minutes what the arithmetic above can only bracket to a factor of five, and the number decides how many organisations should be rotating credentials rather than wondering.
  • Whether the pi adapter's efficiency claim survives a different harness. The gain is reported under the harness it was trained against, which is the one place it is least surprising. The cheap falsifier is the same adapter under Terminus or Claude Code at matched budgets, and the weights are Apache-2.0, so this costs an afternoon.
  • Whether Cerebras or a partner publishes a disaggregated figure against a GPU-only baseline. The 5× is against an aggregated Cerebras configuration, which is the comparison most favourable to the change. Throughput per megawatt at matched tokens per second against an all-GPU deployment is the number an infrastructure buyer actually needs.
  • Still open: Voxtral 3 weights, day 30, with nothing published under mistralai since the NVFP4 quantisation of 15 October; no independent score for Inkling 2 Small, flagged on the 13th; no independent token count behind the Qwen4-Max efficiency claim, asked on the 16th; no four-loop SenseNova-Loop-9B result, asked on the 18th; no n-gram embedding ablation; and no threshold-conditioned fail-open rate from any decision-model vendor, day 41.

Daily, by email

Stay current on AI without the scrolling

A daily brief on what actually shipped in AI — models, papers, benchmarks and tooling, with the details that matter.

Confirmation email first, one message a day, unsubscribe in one click.