Good morning. Here is what matters in AI today, and how to put it to work.
We see a sharp split today: new model and toolchain power on one side, and a cluster of serious supply-chain vulnerabilities on the other.
~3 min read · last 12 hours
In today's issue
01
simple-git allows arbitrary command execution via git.clone() injection
02
simple-git: VISUAL editor env var bypasses unsafe-editor detection
03
GLM 5.3 (753B MoE) lands on Amazon Bedrock for coding and agentic tasks
04
Gemini "Call for Me" may expand to personal contacts, not just businesses
05
Amper becomes the Kotlin Toolchain, unifying the build experience
Main story
simple-git allows arbitrary command execution via git.clone() injection
An OS command-injection flaw in simple-git 3.36.0 (current npm latest) lets any attacker-influenced value passed to customArgs execute arbitrary code, with no include-path block in place.
Why it matters: If your AI pipeline, CI system, or agent scaffolding calls git.clone() with any user-supplied input, you are exposed right now and should audit or pin the package immediately.
What to watch next: Watch for patch adoption velocity in the npm ecosystem: if simple-git 3.36.0 remains the current latest without a forced deprecation, the window for exploitation stays wide open for any app that passes user input to git.clone().
We are tracking a cluster of critical-to-medium advisories across simple-git, vLLM, Docling, and Werkzeug that collectively show how quickly a compromised dependency can become a code-execution or data-exfiltration path in an AI pipeline.
Parameters in GLM 5.3, the new mixture-of-experts model now available on Amazon Bedrock · AWS Machine Learning Blog
Watch · On the feeds
Building a custom website to stay on top of business with ChatGPT
OpenAI
CoreWeave Forge: Demo from Fully Connected 2026, with Corey Sanders
CoreWeave
The Signal
Today's news splits cleanly into two lanes. In the build lane, a heavyweight new model lands on a managed cloud, a major toolchain gets a unified identity, and Google's AI assistant quietly expands into personal calls. In the risk lane, a wave of critical and high-severity advisories in widely used open-source packages (simple-git, vLLM, Docling, Werkzeug) is a reminder that AI-era supply chains inherit every vulnerability in the libraries they depend on. Teams accelerating AI adoption need to be running the same rigorous dependency hygiene on their AI stacks as on any other production code.
All the best, the KYFEX team
“was written by a very tired, overworked, under-slept human being.”
WIRED
Quick hits
Supply-chain vulnerabilities hit AI and dev tooling hard
simple-git: VISUAL editor env var bypasses unsafe-editor detection
A critical flaw means simple-git's safety guard does not classify Git's VISUAL editor fallback as an unsafe editor, leaving a bypass path open for attackers who control environment variables.
Why it matters: Server environments that inherit broad environment variables (common in containerized AI workloads) are the highest-risk targets here.
New model power and toolchain upgrades raise the production bar
A 753-billion-parameter model on a managed cloud, a unified Kotlin build toolchain, and Google quietly extending its AI calling feature into personal contacts all point in the same direction: the raw capability available to production teams keeps rising, but so does the integration surface that needs to be secured and tested.
GLM 5.3 (753B MoE) lands on Amazon Bedrock for coding and agentic tasks
Z.ai's GLM 5.3, a 753-billion-parameter mixture-of-experts model optimized for coding and long-horizon agentic tasks, is now available on Amazon Bedrock with OpenAI-compatible APIs and prompt caching for cost and latency reduction.
Why it matters: Teams already on Bedrock can drop in a frontier-scale coding and agent model without new infrastructure, making it a low-friction option to benchmark against existing model choices.
Gemini "Call for Me" may expand to personal contacts, not just businesses
Google appears to be extending its Gemini Call for Me feature beyond business calls, with APK teardown evidence suggesting it will let the AI send messages or make calls to friends and family on your behalf.
Why it matters: This expansion blurs the line between AI assistant and personal agent, raising consent and impersonation questions that product and policy teams should get ahead of before it ships broadly.
Amper becomes the Kotlin Toolchain, unifying the build experience
JetBrains has rebranded and evolved Amper into the Kotlin Toolchain (0.11 release), consolidating its build tooling under a single, unified entry point for Kotlin projects.
Why it matters: Teams building Kotlin-based AI services or Android AI features should migrate to the Kotlin Toolchain now to stay on the supported, actively developed path.
Systematically improve the accuracy of LLM applications with evaluation, prompting, and memory tuning.
DeepLearning.AI · Free · 1 hour
Put it to work
Try this today
Audit a dependency list for known CVE patterns before merging a PR
You are a security-focused code reviewer. I will paste a list of npm or Python package names and versions. For each package, flag any that are known to have recent command-injection, path-traversal, or cache-poisoning vulnerabilities. For flagged packages, state: (1) the affected version range, (2) the attack surface in plain language, (3) the recommended remediation (upgrade, config change, or removal). Be concise. Here is my dependency list:
[PASTE YOUR package.json or requirements.txt CONTENTS HERE]
Why it helps: With multiple critical advisories hitting widely used packages today, running this check before merging is faster than waiting for your scanner's next scheduled run.
Before you ship it
The risk
AI pipelines that ingest external documents or clone repos on behalf of users inherit every injection and path-traversal flaw in their underlying libraries, and today's advisories show those flaws are actively present in current releases.
Do this
Enforce a locked, audited dependency manifest for every AI service in production, run automated advisory scanning on every build, and treat any package that touches user-supplied file paths or shell commands as high-risk until verified patched.
Ready to ship AI, not just read about it?
KYFEX designs and builds production AI for teams that need it working, not just demoed. Tell us what you're working on and we'll bring the engineering.
