Back Gbhackers Microsoft’s AI Agent Framework Package Compromised on PyPI to Steal Cloud Credentials
Microsoft’s agent-framework Python package was compromised on November 17, 2026, with malicious versions 1.26.1 and 1.26.2 distributing a credential stealer targeting cloud identities, developer secrets, and CI/CD environments.
The incident extends a pattern of AI dependency supply-chain compromises , where trusted software packages become entry points into infrastructure holding sensitive authentication material.
According to the Socket Research Team’s analysis , both releases installed Python startup hooks that collected credentials and transmitted them to an attacker-controlled webhook endpoint. The payloads targeted Azure, Azure DevOps, GitHub Actions, and local configuration stores, combining conventional secret harvesting with requests for fresh cloud identity tokens.
Microsoft’s AI Framework Package Compromised
The attack relies on executable import statements inside .pth files processed during normal Python initialization. This allows the stealer to execute when Python starts in an affected environment, even without an explicit agent-framework import. Similar startup execution appeared in the compromised LiteLLM package , illustrating why package-import monitoring alone can miss malicious activity.
Version 1.26.1 installs agent_framework_meta.pth, which imports agent_framework_meta. Its modified initializer invokes _telemetry._start(), launching credential collection asynchronously.
On Unix, the payload forks and detaches; platforms without fork support use a daemon thread. An _AF_TEL environment flag prevents repeated execution in descendant processes, but fresh process trees can trigger collection again. Errors are suppressed to reduce visible disruption.
Version 1.26.2 relocates execution to af_runtime_diag.pth and af_runtime_diag.py. It restores the original package initializer byte-for-byte, meaning investigators examining only that file could overlook the revised payload. The change also expands collection to GitHub Actions identity endpoints, runner directories, and additional environment-variable patterns associated with workflow authentication.
Both versions inspect environment variables, credential files, and accessible /proc/*/environ entries. Targets include Azure CLI data, Azure DevOps agent credentials, npm and pip configuration, Git credentials, Docker authentication, Hugging Face tokens, and application .env files. This collection resembles other developer credential theft campaigns that turn dependency compromise into broader cloud exposure.
Beyond copying stored secrets, the malware requests Azure managed-identity and federation tokens. Version 1.26.2 additionally requests GitHub Actions OpenID Connect tokens for Azure-related audiences using workflow-provided authentication variables. Consequently, exposure is not limited to environments with long-lived passwords or API keys; it also targets identities available to the running workload.
Collected information is serialized into JSON, compressed with zlib, Base64-encoded, and transmitted through an HTTPS POST to a concealed webhook.site URL with ?t=telemetry. Base64 encoding does not provide encryption. Socket reported no additional payload downloading or code execution in the analyzed collectors.
Organizations should avoid both releases, remove affected installations, and rotate potentially exposed credentials. Investigations should cover developer machines and automation runners, particularly environments resembling those targeted in GitHub Actions credential-stealing attacks .
Responders should inspect both startup-hook variants and outbound telemetry for the listed endpoint. Removing the package alone does not invalidate secrets already stolen or determine whether attackers subsequently abused cloud identities.
Note: IP addresses and domains are intentionally defanged (e.g., [.]) to prevent accidental resolution or hyperlinking. Re-fang only within controlled threat intelligence platforms such as MISP, VirusTotal, or your SIEM.
Give your SOC the alert context on the latest threats to stop incidents quickly -> Integrate Threat Intel from 16K SOCs :
Artificial Intelligence
Cyber security Course
Cyber Security Resources
Cybersecurity
Information Gathering
Information Security Risks
New TheMailer AI Phishing Kit Uses ChatGPT and Claude to Bypass Email Spam Filters
Hackers Target Indian Military Personnel With GravityRAT Spyware Hidden in Fake Job Recruitment Apps
Critical Java4Shell Flaw in OpenJDK Base64 Decoder Enables Remote Code Execution
OPTIMJACK Attack Injects Malicious JavaScript Into CNN, Zoom and Other Major Websites to Steal Data
Hackers Deploy PerfectPioneer RAT and KUKURAT Backdoor to Steal Sensitive Government Data
McDonald’s iOS App Flooded With Suspicious Push Notifications in Possible Cyberattack
The full story
This article is one source in a clustered incident — the cluster page carries the summary, timeline and every other outlet covering it.
