The Agent Skills Directory

[PRIVILEGE_ESCALATION]: The skill implements multiple techniques to bypass container isolation and escalate privileges. This includes using nsenter to join host namespaces, exploiting CAP_SYS_ADMIN for release_agent escapes, and manipulating core_pattern to execute arbitrary code as host root.
[DATA_EXPOSURE_AND_EXFILTRATION]: scripts/escape_enum.sh and the post-escape playbook in references/node-host-pivot.md provide automated methods for harvesting high-value information, including Kubernetes ServiceAccount tokens, kubeconfig files, etcd server keys, and cloud instance metadata (IMDS) from AWS, GCP, and Azure.
[REMOTE_CODE_EXECUTION]: The toolkit includes scripts like scripts/kubelet_exec.py and scripts/runc_cwd_escape.py that facilitate unauthorized command execution across pods and host environments by exploiting runtime working-directory leaks and kubelet API misconfigurations, such as the nodes/proxy GET path.
[PERSISTENCE_MECHANISMS]: Documentation within the skill (references/node-host-pivot.md) details methods for maintaining access to compromised nodes, such as injecting SSH keys into the host's root directory and deploying static pod manifests in /etc/kubernetes/manifests/ to ensure automatic, stealthy execution of attacker code.
[DYNAMIC_EXECUTION]: scripts/nvidiascape_build.sh facilitates an OCI-hook escape (CVE-2025-23266) by compiling a malicious shared object at runtime and building a container image that triggers a privileged host-side NVIDIA hook to load the attacker's library.
[METADATA_POISONING]: The skill's metadata and instructional content promote offensive actions and provide functional exploit code, posing a significant risk if deployed in an environment where the agent has access to containerized or orchestrated workloads.

container-k8s-escape