MAL-2026-20101
Vulnerability from ossf_malicious_packages – Published: 2026-11-17 22:12 – Updated: 2026-11-17 23:12 – Source website| Ecosystem | Package | Versions |
|---|---|---|
| PyPI |
agent-framework
|
1.26.21.26.1
|
Malicious releases of the legitimate "agent-framework" package (Microsoft Agent Framework). Versions 1.26.1 and 1.26.2 add a credential stealer that is not present in the last clean release, 1.26.0. Both releases were published on 2026-11-17 (1.26.1 at 20:35 UTC as a wheel only, 1.26.2 at 21:08 UTC) and have since been removed from PyPI.
How it triggers: each wheel adds an executable .pth startup hook that imports a helper module. Python runs .pth files at every interpreter start, so the code executes as soon as the package is installed, without importing agent_framework. In 1.26.1 the files are agent_framework_meta.pth and agent_framework_meta/_telemetry.py, started from a hook added to agent_framework_meta/init.py. In 1.26.2 they are renamed to afunti_diag.pth and af_runtime_diag.py; agent_framework_meta/init.py is identical to the clean 1.26.0 and the module calls its own _start() on import, so the .pth is the only trigger. The helper detaches into a background process (os.fork, os.setsid), is guarded by the environment marker _AF_TEL, and silences all errors. The rest of the wheel matches 1.26.0 and the dependency is unchanged (agent-framework-core[all]==1.26.0).
What it does (info stealer): it collects hostname, user, working directory and arguments; environment variables whose names contain TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB, HF_, OPENAI, NPM, PIP or CI_; the environment of other processes (/proc/*/environ); credential files (~/.azure, .npmrc, .netrc, .git-credentials, .docker/config.json, Hugging Face token, Azure DevOps agent .credentials, .env files); Azure DevOps OIDC tokens (SYSTEM_ACCESSTOKEN, SYSTEM_OIDCREQUESTURI); and Azure managed-identity tokens from the instance metadata service (169.254.169.254, IDENTITY_ENDPOINT). 1.26.2 adds theft of GitHub Actions OIDC tokens (ACTIONS_ID_TOKEN_REQUEST_URL, ACTIONS_ID_TOKEN_REQUEST_TOKEN, ACTIONS_RUNTIME_TOKEN), adds the /home/runner paths, and raises the limit on captured files from 60 to 80.
Exfiltration: the data is compressed with zlib, base64-encoded and sent by HTTP POST to https://webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb. The URL is stored base64-encoded (_ENC) and is identical in both releases. The User-Agent is af-runtime-diagnostics/1 in 1.26.1 and runtime-diagnostics/1 in 1.26.2.
Wheel sha256: 1.26.1 d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110; 1.26.2 3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9.
Anyone who installed 1.26.1 or 1.26.2 should treat the environment as exposed and rotate the credentials reachable from it, including CI and GitHub Actions secrets. 1.26.0 and earlier are not affected. We have not attributed this to a known campaign.
The wheel ships agent_framework_meta.pth containing import agent_framework_meta, which runs on every Python interpreter start (not just on import agent_framework). The imported module spawns a background thread that harvests installer-side secrets and POSTs a base64+zlib-compressed payload to https://webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb, whose URL is concealed as a base64 literal (_ENC) decoded at runtime. Collection includes: environment variables matching TOKEN/SECRET/PASSWORD/KEY/AZURE/AWS/GITHUB/OPENAI patterns; credential files ~/.azure, ~/.npmrc, ~/.netrc, ~/.git-credentials, ~/.docker/config.json, Hugging Face tokens; /proc/[0-9]*/environ scraping of other processes; Azure IMDS (169.254.169.254) and App Service IDENTITY_ENDPOINT managed-identity token requests across multiple resource audiences (management.azure.com, vault.azure.net, graph.microsoft.com, storage, etc.); and an ADO OIDC token-exchange POST using SYSTEM_ACCESSTOKEN against SYSTEM_OIDCREQUESTURI. Package metadata impersonates Microsoft and mimics the real microsoft/agent-framework README as cover. The name resembles the legitimate Microsoft package.
Versions 1.26.1, 1.26.2 were compromised.
Compromised versions use PTH file to trigger malicious actions on every Python process startup. It then exfiltrates environment variables, credential files and tokens from cloud metadata endpoints.
Category: MALICIOUS - The campaign has clearly malicious intent, like infostealers.
Campaign: 2026-11-compr-agent-framework
Reasons (based on the campaign):
-
exfiltration-env-variables
-
obfuscation
-
exfiltration-cloud-tokens
-
exfiltration-credentials
-
compromised-package
-
abuses-pth
| Type | Indicator | Hashes |
|---|---|---|
| Domain |
webhook.site
|
|
| IP |
169.254.169.254
|
|
| URL |
https://webhook.site/
|
|
| Evidence file |
agent_framework_meta/_telemetry.py
|
sha256
39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663tlsh
f0d17592a94038e3d206726ccc24e61c233bfd4b2c515434becc5b594f8e5bad5b6aad |
| Evidence file |
agent_framework_meta.pth
|
sha256
21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e |
- CWE-506 - Embedded Malicious Code
{
"affected": [
{
"database_specific": {
"cwes": [
{
"cweId": "CWE-506",
"description": "The product contains code that appears to be malicious in nature.",
"name": "Embedded Malicious Code"
}
],
"indicators": {
"evidence_files": [
{
"artifacts": [
"agent_framework-1.26.1-py3-none-any.whl"
],
"path": "agent_framework_meta/_telemetry.py",
"sha256": "39fa329f2a93f734b0bb8f5adb203ea3331d1ae479832a5015a54cb17aabd663",
"tlsh": "f0d17592a94038e3d206726ccc24e61c233bfd4b2c515434becc5b594f8e5bad5b6aad"
},
{
"artifacts": [
"agent_framework-1.26.1-py3-none-any.whl"
],
"path": "agent_framework_meta.pth",
"sha256": "21f28e8517a10ffe431f8d68f5875c85eca4001ae6d118322b0a41873998eb5e"
}
]
},
"iocs": {
"domains": [
"webhook.site"
],
"files": [],
"ips": [
"169.254.169.254"
],
"urls": [
"https://webhook.site/"
]
}
},
"package": {
"ecosystem": "PyPI",
"name": "agent-framework"
},
"versions": [
"1.26.2",
"1.26.1"
]
}
],
"credits": [
{
"contact": [
"inspector-research@amazon.com"
],
"name": "Amazon Inspector",
"type": "FINDER"
},
{
"contact": [
"https://github.com/kam193",
"https://bad-packages.kam193.eu/"
],
"name": "Kamil Ma\u0144kowski (kam193)",
"type": "ANALYST"
},
{
"name": "OX Security",
"type": "FINDER"
}
],
"database_specific": {
"malicious-packages-origins": [
{
"id": "IN-MAL-2026-022818",
"import_time": "2026-11-17T23:10:33.787225233Z",
"modified_time": "2026-11-17T23:04:06Z",
"sha256": "72f7549a69d5529aca2fb787d2fbb62f94e73afd9ca1a832608687e1d185e594",
"source": "amazon-inspector",
"versions": [
"1.26.1"
]
},
{
"id": "pypi/2026-11-compr-agent-framework/agent-framework",
"import_time": "2026-11-17T23:10:36.477156957Z",
"modified_time": "2026-11-17T22:12:19Z",
"sha256": "7170b55e6541c89d914a388534b7ccaeb303abb46def27bda667f5ad2051ed5d",
"source": "kam193",
"versions": [
"1.26.1",
"1.26.2"
]
}
]
},
"details": "Malicious releases of the legitimate \"agent-framework\" package (Microsoft Agent Framework). Versions 1.26.1 and 1.26.2 add a credential stealer that is not present in the last clean release, 1.26.0. Both releases were published on 2026-11-17 (1.26.1 at 20:35 UTC as a wheel only, 1.26.2 at 21:08 UTC) and have since been removed from PyPI.\n\nHow it triggers: each wheel adds an executable .pth startup hook that imports a helper module. Python runs .pth files at every interpreter start, so the code executes as soon as the package is installed, without importing agent_framework. In 1.26.1 the files are agent_framework_meta.pth and agent_framework_meta/_telemetry.py, started from a hook added to agent_framework_meta/__init__.py. In 1.26.2 they are renamed to af_runtime_diag.pth and af_runtime_diag.py; agent_framework_meta/__init__.py is identical to the clean 1.26.0 and the module calls its own _start() on import, so the .pth is the only trigger. The helper detaches into a background process (os.fork, os.setsid), is guarded by the environment marker _AF_TEL, and silences all errors. The rest of the wheel matches 1.26.0 and the dependency is unchanged (agent-framework-core[all]==1.26.0).\n\nWhat it does (info stealer): it collects hostname, user, working directory and arguments; environment variables whose names contain TOKEN, SECRET, PASSWORD, KEY, AZURE, AWS, GITHUB, HF_, OPENAI, NPM, PIP or CI_; the environment of other processes (/proc/*/environ); credential files (~/.azure, .npmrc, .netrc, .git-credentials, .docker/config.json, Hugging Face token, Azure DevOps agent .credentials, .env files); Azure DevOps OIDC tokens (SYSTEM_ACCESSTOKEN, SYSTEM_OIDCREQUESTURI); and Azure managed-identity tokens from the instance metadata service (169.254.169.254, IDENTITY_ENDPOINT). 1.26.2 adds theft of GitHub Actions OIDC tokens (ACTIONS_ID_TOKEN_REQUEST_URL, ACTIONS_ID_TOKEN_REQUEST_TOKEN, ACTIONS_RUNTIME_TOKEN), adds the /home/runner paths, and raises the limit on captured files from 60 to 80.\n\nExfiltration: the data is compressed with zlib, base64-encoded and sent by HTTP POST to https://webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb. The URL is stored base64-encoded (_ENC) and is identical in both releases. The User-Agent is af-runtime-diagnostics/1 in 1.26.1 and runtime-diagnostics/1 in 1.26.2.\n\nWheel sha256: 1.26.1 d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110; 1.26.2 3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9.\n\nAnyone who installed 1.26.1 or 1.26.2 should treat the environment as exposed and rotate the credentials reachable from it, including CI and GitHub Actions secrets. 1.26.0 and earlier are not affected. We have not attributed this to a known campaign.\n\n---\n_-= Per source details. Do not edit below this line.=-_\n\n## Source: amazon-inspector (72f7549a69d5529aca2fb787d2fbb62f94e73afd9ca1a832608687e1d185e594)\nThe wheel ships agent_framework_meta.pth containing `import agent_framework_meta`, which runs on every Python interpreter start (not just on `import agent_framework`). The imported module spawns a background thread that harvests installer-side secrets and POSTs a base64+zlib-compressed payload to https://webhook.site/d381fd80-ec62-42ad-8f5c-d1b15103c9bb, whose URL is concealed as a base64 literal (_ENC) decoded at runtime. Collection includes: environment variables matching TOKEN/SECRET/PASSWORD/KEY/AZURE/AWS/GITHUB/OPENAI patterns; credential files ~/.azure, ~/.npmrc, ~/.netrc, ~/.git-credentials, ~/.docker/config.json, Hugging Face tokens; /proc/[0-9]*/environ scraping of other processes; Azure IMDS (169.254.169.254) and App Service IDENTITY_ENDPOINT managed-identity token requests across multiple resource audiences (management.azure.com, vault.azure.net, graph.microsoft.com, storage, etc.); and an ADO OIDC token-exchange POST using SYSTEM_ACCESSTOKEN against SYSTEM_OIDCREQUESTURI. Package metadata impersonates Microsoft and mimics the real microsoft/agent-framework README as cover. The name resembles the legitimate Microsoft package.\n\n## Source: kam193 (7170b55e6541c89d914a388534b7ccaeb303abb46def27bda667f5ad2051ed5d)\nVersions 1.26.1, 1.26.2 were compromised.\n\n\nCompromised versions use PTH file to trigger malicious actions on every Python process startup. It then exfiltrates environment variables, credential files and tokens from cloud metadata endpoints.\n\n\n---\n\nCategory: MALICIOUS - The campaign has clearly malicious intent, like infostealers.\n\n\nCampaign: 2026-11-compr-agent-framework\n\n\nReasons (based on the campaign):\n\n\n - exfiltration-env-variables\n\n\n - obfuscation\n\n\n - exfiltration-cloud-tokens\n\n\n - exfiltration-credentials\n\n\n - compromised-package\n\n\n - abuses-pth\n",
"id": "MAL-2026-20101",
"modified": "2026-11-17T23:12:16.911370636Z",
"published": "2026-11-17T22:12:19Z",
"references": [
{
"type": "PACKAGE",
"url": "https://pypi.org/project/agent-framework/1.26.2/"
},
{
"type": "PACKAGE",
"url": "https://pypi.org/project/agent-framework/1.26.1/"
},
{
"type": "WEB",
"url": "https://opensourcemalware.com/blog/agent-framework-pypi-compromised"
},
{
"type": "WEB",
"url": "https://bad-packages.kam193.eu/pypi/campaign/2026-11-compr-agent-framework"
}
],
"schema_version": "1.7.4",
"summary": "Malicious code in agent-framework (PyPI)"
}
Drag nodes to arrange them. Double-click a related vulnerability to open it, or a source to visit its link. The Physics button on the left tunes the layout and the simulation; the legend at the bottom left hides or shows nodes by kind and by sighting type.
Sightings
| Author | Source | Type | Date | Other |
|---|
Nomenclature
- Seen: The vulnerability was mentioned, discussed, or observed by the user.
- Confirmed: The vulnerability has been validated from an analyst's perspective.
- Published Proof of Concept: A public proof of concept is available for this vulnerability.
- Exploited: The vulnerability was observed as exploited by the user who reported the sighting.
- Patched: The vulnerability was observed as successfully patched by the user who reported the sighting.
- Not exploited: The vulnerability was not observed as exploited by the user who reported the sighting.
- Not confirmed: The user expressed doubt about the validity of the vulnerability.
- Not patched: The vulnerability was not observed as successfully patched by the user who reported the sighting.
Vulnerability timeline
Sightings and recorded dates, grouped by day · UTC
How durations are calculated
Following the VTEM framework (Black Hat USA 2015), time to patch starts when the vendor or a coordinator was informed, total time to patch at discovery, and disclosure to patch at public disclosure; vendor response time runs from the report to its acknowledgement. Time to exploit and time unpatched start at the earliest discovery or report, falling back to public disclosure. Unpatched time ends at the first recorded patch, or the calculation time when no patch is recorded. Exploitation to patch shows the interval between those two events; a published proof of concept is a public exploit, not exploitation.
Each result shows its source dates. A negative interval means the end event came first; when date-only inputs leave the order open, the result says so. Date-only inputs are approximate. The earliest evidence wins: KEV report dates and sightings are labelled as such. Records describing another vulnerability are listed but never calculated from, and a PoC alone does not establish exploitation. Missing patch evidence does not prove that no patch exists. Filters below change the event list only.
Loading timeline…
The approach is described in our paper Mapping CVEs to MITRE ATT&CK Techniques: A Curated Gold-Set Classifier and the Limits of LLM-Assisted Label Expansion.
Browse all ATT&CK techniques and the vulnerabilities related to each.
Related by attack behaviour
Vulnerabilities whose description is nearest to this one in the vector space of the CIRCL/vulnerability-attack-technique-biencoder model. This is a similarity search over the bi-encoder space (plain cosine), not a classification, and it has no measured accuracy.