Executive Summary
On November 17, 2026, attackers compromised Microsoft's official Agent Framework Python package on PyPI, publishing two malicious versions (1.26.1 and 1.26.2) that automatically executed a credential stealer during Python startup—even without an explicit import of the package. The malware harvested secrets from environment variables, configuration files, and running processes, targeting Azure, AWS, GitHub, Docker, Hugging Face, and OpenAI credentials, as well as Azure managed identity tokens and GitHub Actions OIDC tokens. Both versions exfiltrated stolen data to an attacker-controlled webhook endpoint. PyPI quarantined each release, but the attacker published a second malicious version within 2.5 minutes, indicating persistent publish access. With roughly 200,000 monthly downloads, the blast radius across developer and CI/CD environments was significant.
This incident underscores a deepening threat to AI and cloud-native development toolchains. Supply-chain attacks targeting popular AI frameworks are accelerating as adversaries recognize that developer machines and CI/CD pipelines hold privileged access to multi-cloud environments. Organizations building with open-source AI toolkits must treat package integrity verification and runtime egress monitoring as non-negotiable security controls.
Why This Matters Now
AI agent frameworks are among the fastest-growing dependency categories in enterprise software development, making them high-value supply-chain targets. As agentic AI workflows gain privileged access to cloud infrastructure and secrets, a single compromised package can silently exfiltrate credentials across entire multi-cloud estates before detection—raising the urgency for runtime egress controls and software supply-chain verification.
Attack Path Analysis
Attackers gained unauthorized publishing access to Microsoft's official Agent Framework PyPI account and injected a credential-stealing payload into versions 1.26.1 and 1.26.2, which executed automatically at Python startup to harvest cloud credentials from environment variables, config files, and running processes. The malware silently collected Azure, AWS, GitHub, Docker, Hugging Face, and OpenAI secrets — including OIDC tokens and managed identity access tokens — and exfiltrated them to an attacker-controlled webhook endpoint, potentially enabling downstream cloud resource takeover and persistent unauthorized access.
Kill Chain Progression
This analysis maps confirmed threat intelligence to the full cloud kill chain to show where defensive gaps would emerge as an attack progresses.
Initial Compromise
Description
The attacker gained publishing credentials or session access to Microsoft's official PyPI account for the agent-framework package, likely through credential theft, token hijacking, or compromise of a maintainer's CI/CD pipeline or development environment. Two malicious versions (1.26.1 and 1.26.2) were published on November 17, 2026, embedding a .pth-based payload that auto-executed during Python startup without requiring an explicit import of the package.
MITRE ATT&CK® Techniques
Supply Chain Compromise: Compromise Software Dependencies and Development Tools
Command and Scripting Interpreter: Python
Event Triggered Execution: PowerShell Profile
Unsecured Credentials: Credentials In Files
Unsecured Credentials: Container API
Steal Application Access Token
Exfiltration Over C2 Channel
Masquerading: Invalid Code Signature
Potential Compliance Exposure
Mapping incident impact across multiple compliance frameworks.
NIST SP 800-53 Rev. 5 – Provenance and Integrity of Third-Party Components
Control ID: SR-4
PCI DSS 4.0 – Software and System Components Are Protected from Known Vulnerabilities
Control ID: Requirement 6.3.3
NYDFS 23 NYCRR 500 – Monitoring and Filtering of Privileged Access and Third-Party Software
Control ID: Section 500.14(b)
DORA (EU) 2022/2554 – Management of ICT Third-Party Service Provider Risk
Control ID: Article 28 — ICT Third-Party Risk Management
NIS2 Directive (EU) 2022/0383 – Security of the Supply Chain Including Software Suppliers
Control ID: Article 21(2)(d) — Supply Chain Security
CISA Zero Trust Maturity Model 2.0 – Continuous Validation of Application Integrity in the Software Supply Chain
Control ID: Applications Pillar — Application Security Testing and Software Supply Chain
Sector Implications
Industry-specific impact of the vulnerabilities, including operational, regulatory, and cloud security risks.
Computer Software/Engineering
Developers using Microsoft's Agent Framework PyPI package face direct supply-chain infostealer exposure, with Azure, GitHub, and OpenAI credentials silently exfiltrated during Python startup via compromised dependency injection.
Information Technology/IT
IT teams relying on PyPI-sourced AI tooling risk cloud credential theft across Azure, AWS, and Docker environments, with egress security gaps enabling silent exfiltration to attacker-controlled webhook endpoints.
Financial Services
Financial organizations using AI agent frameworks in cloud-native workflows face PCI 4.0 and NIST.800-53 compliance violations if stolen Azure managed identity or OIDC tokens enable unauthorized lateral movement into regulated environments.
Health Care / Life Sciences
Healthcare AI development pipelines consuming compromised PyPI packages risk HIPAA.164.312 violations, as stolen cloud credentials could expose protected health data stored in Azure or AWS cloud infrastructure.
Sources
- Microsoft’s Agent Framework PyPI package compromised to steal credentialshttps://www.bleepingcomputer.com/news/security/microsofts-agent-framework-pypi-package-compromised-to-steal-credentials/Verified
- agent-framework PyPI Package Compromised with Credential Stealerhttps://socket.dev/blog/agent-framework-pypi-package-compromised-with-credential-stealerVerified
- Microsoft Agent Framework PyPI Package Compromised - StepSecurity Bloghttps://www.stepsecurity.io/blog/microsoft-agent-framework-pypi-package-compromisedVerified
- Breaking: Hijacked Releases of Microsoft's Official Agent Framework Steals Azure and CI Credentialshttps://www.ox.security/blog/breaking-hijacked-releases-of-microsofts-official-agent-framework-steals-azure-and-ci-credentials/Verified
- PyPI Statistics - agent-frameworkhttps://pypistats.org/packages/agent-frameworkVerified
Frequently Asked Questions
Cloud Native Security Fabric Mitigations and ControlsCNSF
Based on the attack progression modeled above, these are the defensive controls that would constrain each stage.
This incident demonstrates how a malicious PyPI package silently harvested cloud credentials from developer and CI/CD environments and exfiltrated them over outbound HTTPS, exposing Azure, AWS, and GitHub identities to attacker-controlled infrastructure. Aviatrix Zero Trust CNSF would likely reduce the blast radius by constraining workload reachability, limiting cross-cloud lateral paths, and restricting unauthorized egress to external webhook endpoints.
Control: Cloud Native Security Fabric (CNSF)
Mitigation: Enforcing workload-aware segmentation across CI/CD and developer environments would likely constrain the execution reach of the malicious package, reducing the number of environments where the payload could silently activate and harvest credentials.
Control: Zero Trust Segmentation
Mitigation: Zero Trust segmentation would likely reduce the scope of privilege escalation by constraining which workloads can interact with identity endpoints and token-issuing services, limiting how far harvested tokens could be leveraged across the cloud environment.
Control: East-West Traffic Security
Mitigation: East-west traffic enforcement would likely constrain the attacker's ability to move laterally across cloud workloads and service boundaries, reducing reachability between environments that hold distinct cloud identities and limiting how many downstream services could be accessed using stolen credentials.
Control: Multicloud Visibility & Control
Mitigation: Multicloud visibility and control would likely surface anomalous outbound traffic patterns from CI/CD and developer workloads to external webhook destinations, reducing the attacker's ability to maintain an undetected C2 channel across cloud network boundaries.
Control: Egress Security & Policy Enforcement
Mitigation: Egress security policy enforcement would likely reduce the attacker's ability to successfully exfiltrate harvested credentials by constraining outbound HTTPS connections from workloads to unapproved external destinations, potentially limiting completed data transfers to attacker-controlled infrastructure.
Even where credential harvesting may have occurred, the constrained lateral paths and restricted egress enforced upstream would likely reduce the attacker's ability to operationalize stolen tokens at scale, limiting the number of cloud resources, repositories, and downstream pipelines reachable through the exposed identities.
Impact at a Glance
Affected Business Functions
- CI/CD Pipeline Integrity and Software Build Systems
- Cloud Identity and Access Management (Azure, AWS, GitHub)
- AI/ML Developer Tooling and Agent Workflow Automation
- Secrets Management and Credential Governance
Estimated downtime: 2 days
Estimated loss: N/A
Sensitive credentials exfiltrated from compromised developer and CI environments include: Azure service principal and managed identity access tokens, Azure DevOps OIDC tokens, AWS access keys and secrets, GitHub personal access tokens and Actions OIDC tokens, Docker registry credentials, Hugging Face API keys, OpenAI API keys, and any additional secrets stored in environment variables or local configuration files on affected developer workstations and CI runner nodes. Approximately 200,000 monthly PyPI downloads indicates a broad potential victim pool across the global developer community. Full scope of successful exfiltrations versus blocked attempts depends on egress log analysis against the attacker-controlled webhook endpoint.
Recommended Actions
Key Takeaways & Next Steps
- • Apply CNSF Egress Security & Policy Enforcement with FQDN-based filtering to block outbound connections from developer and CI/CD environments to untrusted webhook domains such as webhook.site, ensuring stolen credentials cannot be exfiltrated over HTTPS to attacker-controlled endpoints.
- • Deploy Cloud Native Security Fabric (CNSF) inline enforcement and anomaly detection on agentic AI and developer pipeline environments to detect unauthorized Python startup execution patterns, abnormal environment variable reads, and token harvesting behaviors associated with malicious .pth payloads.
- • Enforce Zero Trust Segmentation with least-privilege identity-based policies to ensure that managed identity tokens and OIDC tokens available in CI/CD runners and developer environments are scoped to minimal required permissions, limiting the blast radius of stolen credentials.
- • Implement Multicloud Visibility & Control with centralized traffic observability across Azure, AWS, and GitHub Actions pipelines to surface anomalous cross-cloud API calls, unexpected credential usage patterns, and lateral movement attempts using harvested tokens after initial compromise.
- • Use Egress Security controls with Cloud Firewall (ACF) URL filtering to enforce allowlist-based outbound connectivity for all developer workstations and CI/CD agents, preventing malware embedded in PyPI packages from reaching attacker-controlled HTTPS endpoints and providing log evidence to distinguish blocked from completed credential exfiltration attempts.



