KDCube

Agent Framework Poisoned on PyPI as Ignite Reinvents RAG

Attackers slipped credential-stealing builds of Microsoft's agent-framework onto PyPI on Nov 17, running a secret-grabber at Python startup before the packages were quarantined — rebuild from 1.26.0 and rotate keys. Meanwhile at Ignite 2026, Microsoft recast RAG as agentic retrieval with Foundry IQ...

Highlights

  • Attackers pushed credential-stealing builds of Microsoft's agent-framework to PyPI on Nov 17 — a package pulling ~200K downloads a month — ironically mid-Ignite (BleepingComputer)
  • At Ignite 2026, Microsoft reframed RAG as "agentic retrieval" with Foundry IQ, a single serverless grounding API with query planning, iterative search, and reflection (Microsoft Tech Community)
  • Foundry moved agent ops to GA: Agent Optimizer, multi-turn evaluation, a model router drawing on the top ~50 models, and agents running on Azure Kubernetes Service (Microsoft)
  • OpenAI's GPT-6.2 Astra and Sol shipped Nov 17 to the API and Codex — Astra 20% cheaper than GPT-6 at $8/$40 per million tokens (OpenAI Community)

Key Signals

  1. Credential stealer slips into Microsoft's agent framework — Nov 17, 2026

    Two malicious releases (1.26.1, 1.26.2) of Microsoft's agent-framework ran a stealer at Python startup, harvesting Azure, AWS, GitHub, Docker, Hugging Face, and OpenAI secrets — with 1.26.2 adding GitHub Actions OIDC token theft — before PyPI quarantined both within minutes. Socket, StepSecurity, and Ox Security flagged it; builders on the bad versions should rebuild from 1.26.0 and rotate every exposed credential (BleepingComputer, SecurityWeek).

  2. Foundry IQ turns RAG into a grounding service — Ignite 2026, Nov 17–20

    Foundry IQ (public preview) collapses the ingest→index→retrieve→synthesize pipeline into one permission-aware grounding API powered by Azure AI Search, and now ships in every new Foundry project on a usage-based plan. It sits inside the broader Microsoft IQ context layer (Work/Fabric/Foundry/Web IQ) with 350+ connectors — a direct bid to make hand-rolled RAG stacks optional (Microsoft Tech Community, Microsoft).

  3. Agent operations and model routing go production — Ignite 2026

    Foundry made Agent Optimizer, multi-turn evaluation, and production insights generally available, expanded its model router across the top ~50 models, and previewed Foundry agents on AKS with the Azure SRE Agent and confidential computing. Combined with GPT-6.2 landing in Foundry the same week, the serving and ops surface tightened fast (Microsoft).

Why It Matters / What To Watch

  1. Your agent framework is now a supply-chain target.
    • Pin and hash-verify agent SDK versions; treat a framework bump like a dependency with code-execution rights at import time (BleepingComputer).
    • Scope CI/CD and cloud credentials so a startup-time stealer can't reach OIDC tokens or long-lived keys (SecurityWeek).
  2. Managed grounding changes the build-vs-buy math for RAG.
    • Evaluate Foundry IQ's agentic retrieval against your custom pipeline on permission-respecting recall and latency before committing more engineering to bespoke RAG (Microsoft Tech Community).
    • Watch whether Agent Optimizer's live-signal "continuous improvement loop" actually moves eval scores in production, not just demos (Microsoft).

Quick Links