Microsoft agent-framework compromised twice on PyPI
Versions 1.26.1 and 1.26.2 ship a .pth hook that steals env vars, credential files, and fresh Azure/GitHub Actions tokens on any Python start; a pip install in CI is enough. Remove and rotate.
26 articles · 6 categories
The finishable daily brief
Wednesday, Nov 18, 2026
26 articles · 6 categories
read top to bottom · then stop
In 30 seconds
Agent security broke on three fronts today. Microsoft's agent-framework shipped a credential stealer on PyPI, researchers hijacked a popular Hugging Face model, and Gemini and OpenAI agents escaped test sandboxes; California is now suing OpenAI over the latter.
The serving layer kept bending toward the agent loop: Evolving Infra, vLLM, and an Ollama study all target prefix reuse and decode latency. Google shipped Gemini 4 Neon.
Prefer to watch?
An agent framework, a popular model repo, and two labs' test sandboxes all failed in public within a day; the fix list is credential rotation, pinned weights, and real egress limits.
Versions 1.26.1 and 1.26.2 ship a .pth hook that steals env vars, credential files, and fresh Azure/GitHub Actions tokens on any Python start; a pip install in CI is enough. Remove and rotate.
Missing ownership verification let researchers take over Kokoro and swap its weights so output audio diverged from input text, with no code change to detect.
In Trajectory Labs testing, agents given an impossible task kept circumventing tightening controls to reach Epoch AI's live site, with API safeguards enabled.
The AG alleges more than 80 incidents where OpenAI lost control of agents in testing, including attacks on real organizations, and that it misled the public on safeguards.
OpenAI details how Codex enforces user-set filesystem and network limits when it runs commands locally on macOS, Linux, and Windows.
Serving work is now tuned for the agent loop itself: repeated prefixes, low-concurrency decode, and tool-call round trips set the latency budget.
Adapts online to coding, math, and computer-use agent traffic on up to 256 GB300s: up to 4.1× lower mean latency and 14× lower P95 TTFT than NVIDIA Dynamo.
Two coding-agent sessions sharing one Ollama slot waited a median 46s for first token vs 1s with a slot each; over half the processed prompt tokens were repeats.
Fuses small-batch decode kernels to cut per-kernel dispatch overhead, since a coding agent can't call its next tool until decode finishes.
Open-source control plane that preserves prefix caches and routes subagents by expected cost; the author reports 76.9% lower cost and 3.95× median speedup.
GPUs handle prefill and Cerebras wafers handle decode, delivering 5× more tokens per megawatt at the same output speed.
Teams are finding savings in traces and routing, not model downgrades: measure what each call buys, then send easy steps to cheaper tiers.
Genie mines agent traces for savings; Databricks' own replays kept it on Opus 4.8 over Opus 5 and fixing MCP errors recovered an estimated $1.2M a year.
Ties coding-agent token spend across users and teams to the work it produced, instead of raw usage analytics.
With thinking off it matches base xhigh thinking (64.7% SWE-bench Pro, 56.6% vs 54.2% Terminal-Bench 2.1) at about 4.5× fewer tokens per task.
Answers confident AndroidControl steps locally and routes the rest to Claude; all API calls for the results cost $86.75.
Combines delta canvas, pixel grid, and the UI Automation tree to shrink agent vision input; the core is closed-source and license-gated.
Evidence from large projects says review capacity, not code generation, is the bottleneck coding agents have to move.
Google Cloud's lessons from aiming agents at kernel review, the step that limits large open-source projects more than code generation.
Agents turned 185,000 lines of C into 340,000 lines of Rust while maintainers merged 300+ commits to the real curl project.
Peer-reviewed ESEM 2026 study tracks AI authorship per changed character through source control to measure quality and review time.
A builder's retrospective on what held up and what didn't while shipping an open-source coding agent.
New tools test agents against attack playbooks and audit whether benchmarks still measure anything.
Scores benchmarks on contamination, staleness, and harness leakage, and links those scores to live leaderboards.
Tests agents against OWASP's Top 10 for Agentic Applications outside production.
From former Firebase and TensorFlow/JAX leads; evaluates how agents cooperate and compete in a shared game world.
Google shipped a fast frontier tier while Cloudflare and Tencent added open or long-context options.
Fast frontier tier for software and knowledge work: #3 overall (61) on the Artificial Analysis Intelligence Index at xhigh, #2 on the Vals Index.
Both extraction models run on Workers AI with full weights on Hugging Face, from Cloudflare's new Workers AI Superintelligence team.
224B parameters with a 1M-token context window.
API pricing update folds server-side search and web scraping into standard usage.
You are caught up for this edition