Threat
Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines
First reported wiz.io
Page published · Page updated
Earliest dated coverage: 18 Nov 2026 · First observed: 18 Nov 2026 · Latest dated coverage: 18 Nov 2026
Coverage timeline
Single-source incident — one report is available.
Why it matters
Compromised Microsoft Agent Framework releases expose credentials throughout affected Python environments, making installation checks and credential-exposure assessment important even where the package was never imported.
Attackers published malicious versions 1.26.1 and 1.26.2 of Microsoft's agent-framework PyPI package, using .pth startup hooks to execute a credential stealer whenever Python starts, even without importing the package. The payload collects cloud and CI credentials, including Azure tokens, and sends them to webhook[.]site. Both releases were removed from PyPI; Wiz identifies 1.26.0 as the latest safe version.
Summary
On November 17, 2026, malicious versions of Microsoft's agent-framework PyPI meta-package were published in two successive releases: 1.26.1 at 20:35 UTC and 1.26.2 at approximately 21:08 UTC. Wiz reports that both were unavailable by approximately 21:30 UTC. The latest safe version identified in that reporting is 1.26.0, published five days earlier.[17]
The malicious packages install Python startup hooks that launch a credential stealer whenever Python starts in the affected environment, without requiring an Agent Framework import. The payload targets developer and CI secrets, Azure CLI credentials, Azure DevOps identities, and Azure managed identities. Its collection extends beyond existing secrets: it can ask available identity services to issue fresh tokens. The news report additionally describes GitHub Actions OIDC-token harvesting in version 1.26.2.[0][17]
The stealer is configured to package collected information into compressed, Base64-encoded JSON and send it by HTTPS POST to webhook.site. The supplied evidence describes malicious releases and payload capabilities, but does not establish how many environments installed them or how many exfiltration attempts succeeded. No named actor is attributed to this incident. Publishing access remains unexplained; Wiz's release-artifact analysis suggests a possible upload outside the normal release process, not a confirmed repository compromise.[0][17]
The immediate defensive priorities are to identify and isolate affected hosts and runners, restore clean environments using the identified safe version, investigate outbound traffic to the configured destination, and rotate exposed credentials from a clean system. Azure identity and activity reviews are necessary alongside stored-secret rotation because the payload can obtain newly issued tokens. GitHub Actions workflows deploying to Azure also warrant review where the malicious package was present.[0][17]
Attack chain
- Malicious release distribution: The attacker distributed agent-framework 1.26.1 and 1.26.2 through the legitimate PyPI package on November 17, 2026. How publishing access was obtained is unknown. Wiz identifies the wheel-only 1.26.1 release and its unchanged core dependency as evidence suggesting publication outside the normal project release flow.[0][17]
- Installation of interpreter-startup hooks: Installing 1.26.1 places agent_framework_meta.pth in the site-packages root; 1.26.2 instead places af_runtime_diag.pth there. These hooks make subsequent Python launches execute the malicious import logic even when the application does not import Agent Framework.[0][17]
- Background payload launch: In 1.26.1, agent_framework_meta.pth imports agent_framework_meta, whose __init__.py invokes _telemetry._start(). The payload uses a detached session or, where fork is unavailable, a daemon thread. It sets _AF_TEL=1 to avoid repeat execution in child processes and presents itself using runtime-diagnostics terminology.[17]
- Secret collection and identity-token acquisition: The payload gathers matching environment variables, other processes' environment data, credential files, and Azure instance metadata. It can request Azure DevOps OIDC tokens and Azure managed-identity tokens when the necessary execution context is available. Version 1.26.2 also adds GitHub Actions OIDC-token collection. The evidence does not establish a separate chronological order among these collection operations.[0][17]
- Exfiltration: Collected information is assembled into JSON, zlib-compressed, Base64-encoded, and configured for a single HTTPS POST to webhook.site. The news report supplies the request path /d381fd80-ec62-42ad-8f5c-d1b15103c9bb. Wiz reports that the destination is hardcoded and the payload has no fallback channel.[0][17]
Disclosure timeline
| Date | Event |
|---|---|
| 2026-11-12 18:25 UTC | Legitimate agent-framework 1.26.0 was published as a wheel and source distribution alongside agent-framework-core 1.26.0.[17] |
| 2026-11-17 20:35 UTC | Malicious agent-framework 1.26.1 was published to PyPI.[17] |
| 2026-11-17 ~21:08 UTC | Malicious agent-framework 1.26.2 was published, approximately 33 minutes after 1.26.1.[17] |
| 2026-11-17 ~21:30 UTC | Wiz reports that both malicious versions were no longer available on PyPI by this time.[17] |
| 2026-11-18 | The Hacker News published its incident report, describing both releases as quarantined and relaying investigation and recovery recommendations.[0] |
How it works
The entry point is a compromised package release, not a demonstrated flaw in Agent Framework's application functionality. Wiz reports that legitimate releases normally include both a wheel and a source distribution, together with the same agent-framework-core version. Malicious 1.26.1 was wheel-only and retained a dependency on agent-framework-core[all]==1.26.0. The absence of a corresponding repository tag or release and of malicious code in the main branch supports an out-of-band publishing hypothesis, but does not establish how access was obtained.[17]
Execution relies on Python's processing of import statements in site-packages .pth files at interpreter startup. The malicious hook therefore does not depend on a program explicitly importing agent-framework. The news report notes that Python launches associated with pip can also trigger the stealer in an affected environment.[0][17]
The 1.26.1 loader separates collection from the launching process through a detached session or daemon-thread fallback. Its _AF_TEL=1 environment marker prevents repeat execution in child processes. Names such as _telemetry, the description 'Lightweight runtime diagnostics hook', and the user-agent af-runtime-diagnostics/1 are intended to resemble ordinary SDK telemetry.[17]
Collection uses substring matching against environment-variable names, including TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB, OPENAI, HF_, FOUNDRY, NPM, PIP, NUGET, SYSTEM, BUILD, and AGENT. Wiz also describes reads from up to 300 /proc/*/environ files and up to 60 credential files. Targets include Azure CLI data, .npmrc, pip.conf, .netrc, .git-credentials, Docker configuration, Hugging Face tokens, Azure DevOps agent credentials, and .env or JSON files under /app and /workspace.[17]
When SYSTEM_ACCESSTOKEN and SYSTEM_OIDCREQUESTURI are present, the payload can request an Azure DevOps OIDC token for the AzureADTokenExchange audience. It also requests Azure managed-identity tokens through IMDS or an App Service identity endpoint; the news report identifies IDENTITY_ENDPOINT and MSI_ENDPOINT as endpoint-selection inputs. Wiz describes token requests for 12 resources, including management, Graph, Key Vault, Storage, AI, Search, Container Registry, and Azure DevOps services. These are token-acquisition capabilities, not merely reads of previously stored secrets.[0][17]
Version 1.26.2 changes the startup module to af_runtime_diag.py, broadens credential-related environment-variable matching, and adds GitHub Actions OIDC-token collection. These additions are described in the supplied news report; the supplied Wiz excerpt does not independently document them.[0]
Exfiltration uses a JSON document transformed with zlib compression and Base64 encoding before an HTTPS POST. Both supplied reports identify webhook.site as the destination. The exact request path is supplied only by the news report, while the hardcoded destination and absence of a fallback channel are stated by Wiz.[0][17]
Affected versions and patch status
| Product | Affected | Patch status |
|---|---|---|
| Microsoft Agent Framework: agent-framework PyPI meta-package | Malicious versions 1.26.1 and 1.26.2, published on November 17, 2026. Their startup hooks affect Python launches within environments where the malicious release is installed. | Both releases were reported quarantined or unavailable. Wiz reports unavailability by approximately 21:30 UTC on November 17 and identifies the preceding release, 1.26.0, as the latest safe version. The supplied recovery guidance calls for restoring a clean environment rather than treating package withdrawal as host remediation.[0][17] |
Indicators of Compromise
| Type | Indicator | Context |
|---|---|---|
| domain | webhook.site | Configured exfiltration host. Correlate domain observations with the incident-specific request path, malicious package versions, or startup-hook artifacts rather than treating a domain hit alone as proof of this compromise.[0][17] |
| other | /d381fd80-ec62-42ad-8f5c-d1b15103c9bb | Exfiltration HTTP request path reported by The Hacker News; use with the webhook.site host when searching proxy or HTTP telemetry.[0] |
| file-path | agent_framework_meta.pth | Startup-hook filename installed in the site-packages root by malicious version 1.26.1.[0][17] |
| file-path | af_runtime_diag.pth | Startup-hook filename installed in the site-packages root by malicious version 1.26.2.[0][17] |
| file-path | agent_framework_meta/__init__.py | Version 1.26.1 loader file that calls _telemetry._start() after the startup hook imports agent_framework_meta.[17] |
| file-path | af_runtime_diag.py | Top-level malicious module introduced in version 1.26.2, according to the news report.[0] |
| other | af-runtime-diagnostics/1 | Payload user-agent reported by Wiz, presented as legitimate-looking runtime diagnostics.[17] |
Key takeaways
- This compromise makes Python startup, rather than Agent Framework application use, the critical execution boundary. Installed startup hooks can trigger during unrelated Python launches.[0][17]
- The identity exposure extends beyond credentials already on disk. Azure DevOps federation and managed-identity token requests make identity-log investigation a necessary companion to stored-secret rotation.[0][17]
- Treat the release-flow anomaly as an investigative lead, not a proven root cause. The available evidence suggests an out-of-band publication path but does not establish how publishing access was compromised.[17]
- Do not transfer attribution from the earlier durabletask incident. Wiz explicitly states that it has not linked the two compromises.[17]
Defensive actions
- Identify installations of agent-framework 1.26.1 or 1.26.2 on developer hosts and CI runners, and inspect their site-packages directories for the two reported .pth startup hooks.: Version checks and hook artifacts directly address the malicious releases. An application need not import Agent Framework for the payload to run, so import-based usage inventories alone do not establish whether an environment was exposed.[0][17]
- Prioritize installation-history review around November 17, 2026, from 20:35 to approximately 21:30 UTC, while also checking later execution in environments that retained either malicious release.: The news report recommends a window ending around 21:20 UTC, but Wiz reports both releases unavailable by approximately 21:30 UTC. Neither withdrawal time is a cutoff for execution of a startup hook already installed locally.[0][17]
- Isolate affected hosts and runners, then restore from a clean environment pinned to the identified safe version, 1.26.0.: The supplied response guidance explicitly recommends isolation and clean restoration. PyPI quarantine does not address the locally installed startup hooks that execute during subsequent Python launches.[0][17]
- Search DNS telemetry for webhook.site and proxy or HTTP telemetry for the request path /d381fd80-ec62-42ad-8f5c-d1b15103c9bb; correlate matches with affected package installations and hook artifacts.: The payload is configured to send collected secrets to that host and path. Wiz describes a hardcoded destination without a fallback channel, making destination-specific investigation particularly relevant.[0][17]
- Rotate exposed credentials from a clean system, covering secrets available in environment variables and the credential files targeted by the payload.: The news report explicitly recommends clean-system rotation. The collection scope includes Azure CLI, package-registry, Git, Docker, Hugging Face, Azure DevOps agent, and application configuration credentials, so exposure review should not be limited to Azure secrets alone.[0][17]
- Audit Azure identities and review sign-in and activity logs with attention to identities accessible from affected compute and Azure DevOps pipelines.: The payload can request new managed-identity and federated tokens, not just copy stored credentials. The supplied reporting specifically warns that rotating stored secrets alone does not cover those token-acquisition paths.[0][17]
- Review GitHub Actions workflows that deploy to Azure when their execution environments contained a malicious release, particularly version 1.26.2.: The news report both recommends this audit and identifies GitHub Actions OIDC-token harvesting as an added capability in 1.26.2.[0]