Microsoft's APM: a dependency manager for your AI agent setup
One manifest replays skills, prompts and MCP servers across Claude Code, Cursor and Copilot. Plus an open Qwen fine-tune, GPU partitioning, AI pentest agents.

Copy markdown
apm.yml pins your agent context like package.json
Microsoft open-sourced APM, a package manager that declares your skills, prompts, agents, hooks and MCP servers in one apm.yml, so `apm install` rebuilds the exact same setup for anyone who clones the repo — across Claude Code, Cursor, Copilot, Codex, Gemini, Windsurf and Kiro. It's MIT-licensed, installs with `pip install apm-cli` (Python 3.10+), and ships a lockfile that pins and content-hashes every dependency, plus hidden-Unicode scanning and an apm-policy.yml for org-wide governance.
Nous ships Arete-27B, a de-propagandized open Qwen
Nous Research released Arete-27B, an Apache-2.0 supervised fine-tune of Alibaba's Qwen3.8-27B with open weights on Hugging Face. On the team's own 803-prompt eval it cuts propaganda-tagged answers from 22% to 0.9% and lifts the fact score from 2.32 to 2.98 out of 5, while coding (Terminal-Bench-2.1) edges from 67% to 70% — scores graded by an LLM judge, so treat them as directional.
SageMaker can now slice one GPU into seven
AWS added NVIDIA MIG partitioning to SageMaker Inference: carve one A100 or H100 into up to 7 hardware-isolated slices (4 on G7e, 2 on G7) and run a different model in each via two new API fields. AWS benchmarks an H100 mig-3g.40gb slice at 18–34% lower cost per million output tokens than a whole A100 — but only above a workload-dependent concurrency crossover, and serverless and single-model endpoints are excluded.
AI pentest agents harvested Apple signing certs for ~$1K
Check Point documented ARTEX, an open-source LLM pentest platform now abused in the wild: over two weeks one operator's agents (DeepSeek v4.1 Flash, Kimi K3.1 Pro) hit gray-market iOS signing shops and pulled 300+ Apple certificates — 3 fully valid — for an estimated $700–1,500 in tokens. With 1,608 ARTEX panels publicly reachable, exposed perimeters are cheap to mass-attack: strip hardcoded creds from frontend bundles, enforce auth on every API, and test for IDOR and injection.