The byte-level path is compact. FACT: version 1.26.1 installs agent_framework_meta.pth, which imports agent_framework_meta/_telemetry.py; 1.26.2 uses af_runtime_diag.pth and af_runtime_diag.py. During normal Python startup, .pth processing loads the stealer, which inspects environment variables, /proc/*/environ, credential files, CI identity endpoints, and Azure managed-identity endpoints, then compresses and Base64-encodes the collection and POSTs it to a webhook.site URL. The distinguishing user agents are af-runtime-diagnostics/1 for 1.26.1 and runtime-diagnostics/1 for 1.26.2. Version 1.26.2 additionally collects GitHub Actions OIDC material. Persistence is therefore Python-interpreter persistence, repeatedly firing while the malicious package remains installed—not OS-level autostart or a documented second stage. Elastic StepSecurity OSV
I would separate proof into three layers. Installation: strongest is the malicious .pth and module present in the active environment, paired with package metadata or installation logs and the wheel hash—1.26.1: d27fa34414e6201d5e03e559ef634f020057554d7b3f4898dd5d7cd3e989e110; 1.26.2: 3d20a8bc91351a9a16b6db243b848ceca0206b7a8623b02d2af553a9c9f7e0a9. A lockfile or cached wheel proves exposure/download, not installation. Execution: prioritize Python telemetry showing the malicious module loaded, followed by accesses to /proc/*/environ, credential files, OIDC endpoints, or 169.254.169.254; a matching .pyc is supporting evidence, not conclusive by itself. Exfiltration: the strongest available host-side evidence is a Python-originated POST to the identified webhook.site destination with the matching user agent and nonzero response; DNS alone proves neither payload execution nor data transfer. Without host and network telemetry, package presence cannot establish that a Python interpreter started or that exfiltration succeeded. JFrog Elastic
ASSESSMENT: recreate the affected virtual environment or CI runner from a trusted release regardless; removal alone leaves too much room for a missed copy of the startup hook. A complete OS rebuild is not automatically required because the published analysis found no second stage or OS persistence. It becomes warranted when execution occurred on a shared or long-lived host, integrity cannot be bounded, or telemetry is incomplete. Conversely, if investigators prove the wheel was only downloaded, never installed—or installed but no site-loading Python process ran—host rebuilding and mass credential rotation are disproportionate. Rebuilding also does not invalidate credentials already transmitted.
Rotate only secrets that were readable during the execution window: exposed Azure service-principal secrets, AWS keys, GitHub tokens, Azure DevOps PATs or service-connection credentials, npm/PyPI tokens, Git credentials, Docker registry authentication, Hugging Face/OpenAI keys, and secrets from .env, .npmrc, .netrc, .git-credentials, or ~/.azure. Scope that list using process environments and file-access evidence. GitHub/Azure DevOps OIDC and Azure managed identity require different handling: cancel affected jobs and sessions, review issuance and cloud sign-ins, restrict or temporarily disable federated trust or identity role assignments, and let or force issued tokens to expire; there is no long-lived OIDC or managed-identity “secret” to rotate indiscriminately. Socket OSV