[ctx]ctxseccontext security: attacks on the harness and context layer of AI agents, plus exploited-now news, hunting and playbooks.

← Context & Harness Attacks · markdown

official (GitHub/PyPA advisories GHSA-764m-3ppc-5m73 / PYSEC-2026-4388; Wiz research) · published here

Microsoft's agent-framework package on PyPI shipped two malicious releases (1.26.1, 1.26.2) that stole Azure, CI and GitHub OIDC tokens on every Python start

  • Treat any host, container or CI job that installed 1.26.1 or 1.26.2 as compromised. Rebuild from a clean image pinned to agent-framework==1.26.0.
  • Rotate everything reachable from it: Azure CLI sessions, ADO PATs and service connections, registry tokens, git credentials, and repo/org secrets exposed to the job.
  • Audit Azure: managed-identity and federated-credential sign-ins from affected VMs, App Services, AKS nodes and pipelines; Wiz notes stolen tokens can stay valid up to 24 hours.
  • Agent stacks are a target: pin agent frameworks by hash, use a package cooldown before taking new releases, and don't give CI jobs that install them id-token: write unless they need it.

Sources: GitHub Advisory GHSA-764m-3ppc-5m73: Malicious code in agent-framework (PyPI) · PyPA advisory PYSEC-2026-4388 (OSV) · Wiz: Microsoft agent-framework compromised (Nov 17, 2026)