Security ·
Microsoft Agent Framework: malicious versions 1.26.1 and 1.26.2 on PyPI stole credentials
On 17 November 2026 attackers published two malicious versions of Microsoft's agent-framework package on PyPI. They steal Azure and CI/CD credentials. The last clean version is 1.26.0.
Microsoft Agent Framework is a library companies use to build AI agents in Python. On 17 November 2026 versions 1.26.1 and 1.26.2 of the agent-framework package appeared on PyPI with added credential-stealing code. According to analyses from Wiz, OX Security and others, the malicious code targets Azure environments and CI/CD pipelines, meaning the secrets that developer machines and build servers usually keep in environment variables and config files.
The malicious versions still depended on the legitimate agent-framework-core package at version 1.26.0, so applications kept working normally and nothing looked wrong. OX Security reports that after 1.26.1 was reverted, the attacker uploaded 1.26.2 with the same payload in a different location. Sources name 1.26.0 as the last clean release. Reports differ on how the attacker obtained publishing access, and we found no official Microsoft statement in the sources we could verify.
Does this affect you? If your project, Docker image or CI job installed agent-framework 1.26.1 or 1.26.2, yes. Search lockfiles, build history, virtual environments and container images. Wiz recommends pinning to 1.26.0. OX Security adds that every secret reachable from a host where the package ran should be treated as exposed: cloud keys, GitHub and Azure DevOps tokens, registry credentials and LLM API keys.
The lesson is broader than one library. AI agents usually get wide permissions and run in environments that hold valuable keys, which makes their dependencies an attractive target. A single unpinned dependency that updates itself during a build is enough.
What it means for companies
Check now whether 1.26.1 or 1.26.2 appears in your lockfiles or builds, and if so rotate every secret reachable from those machines. In the long run, pin versions with hashes, give agents only the permissions they need and keep keys away from the environment where packages are installed. When we build custom AI agents we do this from day one, with human approval for steps that have consequences, as described in AI guardrails. A review of dependencies and pipelines is part of cyber security.
Sources: Wiz: Microsoft agent-framework Compromised: Credential Stealer Targets Azure and CI Pipelines, OX Security: Hijacked Releases of Microsoft's Official agent-framework Steal Azure and CI Credentials