Malicious agent-framework wheels hit PyPI on November 17, 2026 with code that steals Azure, GitHub Actions, and CI/CD credentials via Python .pth startup hooks. Socket’s analysis and Wiz’s incident write-up are the technical sources; Microsoft’s last verified release is python-1.26.0 on GitHub from November 13, 2026.
Timeline: what happened on November 17, 2026
| Time (UTC, per vendors) | Event |
|---|---|
| ~20:35 | 1.26.1 wheel published to PyPI |
| ~20:54 | Upwind reports issue on Microsoft’s GitHub tracker |
| ~21:08 | 1.26.2 follows with modified payload |
| ~21:30 | PyPI quarantines both versions |
| Later | Microsoft repo maintainer confirms pull; investigation ongoing |
Unlike a typical dependency confusion typo, agent-framework is the real Microsoft package name with substantial weekly download volume. That makes this a publisher-account or release-pipeline compromise question, not an impersonation package with a misspelled name.
agent-framework PyPI compromise — supply chain link audit illustration
Mechanism: why .pth hooks are nasty
Wiz explains the infection path:
- Malicious wheel drops
agent_framework_meta.pth(1.26.1) oraf_runtime_diag.pth(1.26.2) into site-packages. - Python’s startup loader executes every
.pthline before user code runs. - Hook imports
_telemetrydisguised as diagnostic telemetry (af-runtime-diagnostics/1user-agent string). - Collector forks a detached process (or daemon thread) and scans:
- Environment keys matching TOKEN, SECRET, AZURE, GITHUB, OPENAI, HF_, and similar substrings.
- Up to 60 credential files (Azure CLI JSON,
.npmrc,.docker/config.json,.env, Azure DevOps agent credentials). - Other processes’
/proc/*/environentries (hundreds of PIDs).
- Requests Azure managed identity and Azure DevOps OIDC tokens; 1.26.2 adds GitHub Actions OIDC.
- zlib + base64 payload POSTed once to
webhook[.]site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb.
You do not need import agent_framework for step 2–6 to run—any python CLI invocation in the venv triggers collection.
What Microsoft’s clean release actually ships
The legitimate python-1.26.0 release notes mention routine dependency range bumps (for example Anthropic 2.x support). There is no python-1.26.1 tag in that repository at the time of writing.
Analysts note 1.26.1 was wheel-only and still depended on agent-framework-core==1.26.0, suggesting an upload outside the project’s paired wheel+sdist workflow.
Builders using Microsoft Foundry agent previews or MAI Code harness news should audit any Python worker that installed agent-framework during the exposure window.
Incident response checklist
1. Find exposure
- Grep lockfiles, Dockerfiles, devcontainer specs, and CI logs for
agent-framework==1.26.1or1.26.2. - Search images built November 17, 2026 UTC evening.
2. Confirm execution artifacts
On disk inside affected virtualenvs:
site-packages/agent_framework_meta.pthsite-packages/agent_framework_meta/_telemetry.py(1.26.1)site-packages/af_runtime_diag.pth(1.26.2)
3. Network forensics
Proxy or DNS logs for the webhook.site UUID above. Vendors note HTTP 200 implies acceptance; 410/429 may mean the sink rejected data—but do not treat absence of logs as safety.
4. Rotate secrets
Prioritize:
- Azure service principals and managed identities on the host.
- GitHub Actions OIDC and
ACTIONS_ID_TOKENscopes (1.26.2). - Azure DevOps
SYSTEM_ACCESSTOKENpipelines. - Hugging Face, OpenAI, and npm tokens in developer homedirs named in Socket’s IOC list.
5. Rebuild environments
- Delete venvs; reinstall from
agent-framework==1.26.0with hash pinning. - Ban floating
>=pins on agent-framework until Microsoft publishes hardening guidance.
CI and container rebuild paths after a poisoned Python wheel
How to protect agent pipelines going forward
Supply-chain attacks on agent runtimes are now recurring:
- Compare with Postman shipping Codex plugins by default and GTIG IDE backdoors.
Minimum practices:
| Control | Rationale |
|---|---|
| Hash-pinned PyPI | Blocks retroactive malicious wheels |
| Internal mirror with delay | Lets Socket/Chainguard block lists propagate |
| Ephemeral CI runners | Shrinks secret residence time |
| OIDC audience restriction | Limits token utility if stolen |
Teams running autonomous coding agents should read AgentBeam’s incident companion for detection patterns aimed at agent hosts.
What Socket and Upwind saw in the payload
Socket’s reverse engineering is the most granular public write-up. Highlights that matter for your runbook:
- Disguised telemetry — module names like
_telemetryand user-agentaf-runtime-diagnostics/1mimic legitimate SDK diagnostics. - Substring env harvesting — not a fixed allowlist; any variable containing
TOKEN,SECRET,KEY,AZURE,GITHUB,OPENAI,HF_,FOUNDRY, orARM_may be copied. - Cross-process reads — up to 300
/proc/*/environscans huntingSYSTEM_ACCESSTOKEN,IDENTITY_HEADER, and GitHubACTIONS_ID_TOKENpatterns in 1.26.2. - Azure file globs —
~/.azure/**/*.json,/home/*/.azure, and/root/.azureJSON stores. - Single-shot exfil — one POST, no retry; failures are swallowed, so defenders cannot assume “failed upload means safe.”
Cybersixt’s timeline adds repository color: a Microsoft maintainer commented on GitHub within hours that compromised versions were removed and investigation continues—use that issue thread for official status, not rumor forums.
Why this hit agent teams especially hard
agent-framework is not an obscure typo-squat. It is the Python entry point for Microsoft’s Agent Framework used in samples that pair with Azure AI Foundry, Semantic Kernel-style orchestration, and multi-agent demos.
If your platform team ran pip install -U agent-framework on November 17 evening UTC to chase features, you may have pulled a wheel never tagged python-1.26.1 on GitHub. That mismatch is your smoking gun in SBOM audits:
| Signal | Benign 1.26.0 | Malicious 1.26.1 / 1.26.2 |
|---|---|---|
Git tag python-1.26.x | Present for 1.26.0 | Absent for 1.26.1 |
| PyPI artifact shape | Wheel and sdist with aligned core | Wheel-only per Wiz |
| Startup side effect | No rogue .pth telemetry | .pth executes every Python launch |
Compare with AgentCorruption on Bedrock AgentCore—different cloud, same lesson: agent hosts are secret vaults.
PyPI mirrors and delayed registries
If your org uses delayed PyPI mirrors or private caches, ensure block lists include 1.26.1 and 1.26.2 wheel hashes even after PyPI removal. Socket and Wiz published IOC hashes the same evening; mirrors that do not ingest third-party advisories may still serve quarantined builds until you pin explicitly.
Developer FAQ: pip, poetry, and containers
I use Poetry/pip-tools—am I safe? Only if lockfiles never resolved during the exposure window. Re-lock against 1.26.0 and diff hashes.
I only build containers on Mondays. Check intermediate layers baked after 2026-11-17 20:35 UTC for the bad version strings.
Does uninstalling fix it? Remove the package and delete leftover .pth files; rebuild the venv. Assume secrets already left the host if Python ran once.
Should I ban webhook.site outright? Many legitimate testers use it; for production agents, egress allow lists are safer than domain blocks alone.
Longer-term controls for agent pipelines
- Two-person release on PyPI publish tokens for any package that touches cloud SDKs.
- Sigstore or company signing attestation on wheels before internal mirrors ingest.
- OIDC federation scopes narrowed so a stolen Actions token cannot reach production subscriptions.
- Runtime monitoring on hosts that run unattended
python -mworkers so unexpected child Python processes and egress to paste-bin domains surface quickly.
Pair with the week’s other agent attacks
This compromise landed the same week as:
- FBI ARTEX FLASH verification (offensive agent framework abuse).
- Google GTIG SomberMeme IDE backdoor (developer workstation social engineering).
- Spillway prompt-injection guard launches (defensive tooling).
Treat them as one November cluster: attackers target where agents load code—PyPI, IDE settings, and autonomous pentest consoles.
Microsoft Agent Framework in your stack
If you are standardizing on Microsoft’s agent abstractions, map where agent-framework appears:
| Layer | Typical install path | Risk if poisoned |
|---|---|---|
| Local dev venv | pip install agent-framework[all] | Developer laptop secrets |
| CI test runner | Cached wheels between jobs | GITHUB_TOKEN, Azure federated creds |
| AKS / App Service worker | Container pip install during build | Managed identity tokens |
| Foundry sample notebooks | Tutorial requirements.txt | Subscription-scoped keys |
None of those layers should use floating latest on a package that can invoke cloud control planes. Pin ==1.26.0 until Microsoft ships a signed 1.26.3+ with an explicit compromise postmortem.
Legal and communications notes
Security teams often debate whether to email customers about a transitive dependency. For agent-framework, assume direct intent: attackers chose a high-trust vendor namespace. If you shipped any service whose Dockerfile upgraded Python deps on November 17 UTC evening, issue an internal severity-1 rotation even without external breach evidence.
Vendors such as Ox Security and Wiz published parallel analyses within hours—use them as secondary citations in tickets, but Socket + GitHub release tag remain the technical anchors for engineering.
Hunting queries (conceptual)
Adapt to your SIEM; these are patterns, not vendor signatures:
- Filesystem: new files matching
agent_framework_meta.pthoraf_runtime_diag.pthundersite-packages. - Process: short-lived child
pythonafter parent shell invoked in CI, with_AF_TEL=1in environment snapshots. - Network: HTTPS POST bodies with base64+zlib blobs to webhook.site subdomains from build agents.
- Identity: sudden Azure AD token minting for multiple resources from one runner IP within seconds.
Pair hunting with NIST-NSA-CISA agent control guidance if your program is still immature.
Rotation priority: start with cloud control-plane credentials that can create new users or export storage—those keys monetize fastest for attackers. Developer OpenAI and Hugging Face tokens matter next for LLM abuse, but they rarely grant domain takeover by themselves. Document every rotation ticket with the exact wheel hash you removed so auditors can close the incident.
What this is not
- Not a vulnerability in Microsoft’s agent source code on GitHub 1.26.0—IOC focuses on PyPI contents.
- Not proof that GitHub Copilot or Foundry clouds were breached; impact is environments that installed the bad wheel.
- Not related to FBI ARTEX offensive framework abuse except in the broad sense of November agent threat volume.
