Agent Framework on PyPI: trace the identity exposure
Two malicious Agent Framework releases reached the official PyPI project on November 17. An enterprise response needs to cover environments, caches and the identities the code could access.
Ler em portuguêsBy Pentesty
A brief distribution window can leave lasting copies
Analyses by Aikido and Wiz place the uploads on November 17, 2026, at 20:35 and 21:08 UTC. Wiz reports that neither was available by approximately 21:30 UTC. When checked on November 18, PyPI displayed 1.26.0 as the current version. This timeline does not confine an investigation to the minutes of public distribution.
Dependency caches, internal mirrors and images may retain previously acquired files. A company that resolved dependencies during that period should connect the acquired file to the builds and environments receiving it, including later reruns. Start with installation and resolution records showing timestamps, versions, available hashes and service owners. A reference in a requirements file alone does not demonstrate installation.
Execution did not require importing the framework
According to StepSecurity’s static analysis, the distributions added .pth files to site-packages. When Python processes those files during startup, their import lines can execute code, as documented by the Python project. The first package used agent_framework_meta.pth; the second used af_runtime_diag.pth, with its collector in a different module. Checking only the first version’s filenames leaves a gap.
This does not mean every installation immediately executed the payload, or that every virtual environment on the same computer was affected. Establish which interpreter processed the compromised directory and when. Isolate the environment before inspecting it and use trusted tools: starting its Python or pip can trigger the hook. Rebuild from a clean base while preserving the records needed for investigation.
Keep conclusions within the evidence
In a November 18 update at 14:00 UTC, Wiz reports indications of a compromised certificate and private key for Microsoft’s ESRP service. This is the researchers’ finding; we do not present it as a Microsoft-confirmed cause. Direct access to issue 10411 was unavailable during this investigation. We also do not attribute the uploads to an account solely because it appears as a maintainer.
StepSecurity advises distinguishing a connection to the shared webhook.site domain, a blocked attempt and a completed transfer to the incident’s specific path. Correlate the process, timestamp and HTTP outcome where those records exist. Missing records or insufficient retention leave uncertainty. The project’s monthly download volume is neither the number of malicious installations nor a count of affected companies.
Define recovery through evidence
Before resuming delivery, assemble the inventory of rebuilt environments, blocks on both versions in package controls, and the review of caches and images. Connect each potentially exposed identity to its containment decision and investigated events. Keep exposure assumed for response purposes separate from demonstrated malicious use.
A clean earlier version does not remove the need to review the complete dependency resolution. Our guide to PyPI release-file provenance supports that acceptance decision. For identity closure, follow the approach to verifiable revocation: confirm that previous access has ended, necessary operations work and out-of-scope operations are refused, using authorized tests.
Frequently asked questions
Does using Agent Framework mean my company was affected?
Not automatically. Look specifically for 1.26.1 and 1.26.2 in resolution records, installations, caches and images. Then establish whether Python processed the hooks in that environment and which identities were accessible.
Is it safe to start with pip show or pip uninstall?
Not in the suspected environment before containment. These tools start Python and may process the hook. Preserve evidence with trusted tools and plan reconstruction from a clean base.
Does replacing keys resolve the entire exposure?
Rotation is one part of the response. Also investigate previously issued tokens, managed identities, federated trust and resource activity. Document what was invalidated and what still relies on restrictions or monitoring.
Does a successful build rule out malicious execution?
No. The reports describe background collection while legitimate dependencies remained intact. Connect the build to execution and network evidence, rather than treating functional success as proof of safety.
Sources and references
- Wiz: analysis and November 18 update
- StepSecurity: static analysis and containment
- SafeDep: credential collection and identity boundaries
- Aikido: timeline and release differences
- PyPI: current agent-framework project state
- Python: startup file processing
- BleepingComputer: incident coverage
- The Hacker News: incident coverage
- SecurityWeek: coverage and investigation limits
Original Pentesty analysis, published November 18, 2026. The proposed scenarios are editorial guidance, with scope and validation to be defined for each environment.