T05 · Unauthorized Access and Privilege Escalation
- Location
scripts/install_ubuntu_24_mvp.sh:23- Finding
Persistent Root-Equivalent Access Through Docker Service and Group Membership
- Content
View full analysis
- Remediation
View remediation
Security audit
Security checks for vulnerabilities and agentic risk
This skill is a disclosed EDA workflow helper, but installation and dashboard defaults create review-worthy host-privilege and data-exposure risks.
Install or run this only in a disposable VM or dedicated EDA machine. Review the setup scripts before use, avoid permanent docker-group membership where possible, do not expose the dashboard outside localhost, pin Docker images by digest, and treat specs, RTL, logs, and GDS outputs as potentially sensitive.
scripts/install_ubuntu_24_mvp.sh:23Persistent Root-Equivalent Access Through Docker Service and Group Membership
scripts/serve_multi_project_dashboard.py:27Unauthenticated Network Exposure of EDA Projects and Generated Artifacts
scripts/run_synth.sh:14Yosys Command Injection Through Unvalidated Synthesis Parameters
scripts/bootstrap_eda_demo.sh:86Unverified and Mutable Third-Party EDA Dependencies
The declared description presents a broad orchestration skill for running an open-source EDA flow from specification through GDS and for creating/iterating/auditing AgentSkills. This code chunk does not perform those workflow-driving tasks. It only reads existing project/run/report files under a fixed workspace path and generates a multi-project HTML dashboard plus per-project detail pages. While the dashboard summarizes outputs from an EDA flow and links to generated artifacts like GDS, that is only a monitoring/reporting function, not the declared primary capability of driving the flow end to end. Therefore the description materially overstates and misrepresents this code chunk’s behavior.
The declared description promises a broad workflow orchestration capability for AI-driven chip design, including turning specs into RTL/GDS and auditing AgentSkills. This code chunk does something much narrower and materially different: it hosts previously generated demo artifacts via a web server and refreshes summary files by invoking helper scripts. While report-summary generation is loosely related to EDA outputs, the primary behavior here is demo artifact serving over HTTP, with hardcoded paths to one example design. That network-serving capability and narrow purpose are not reflected in the declared description, so this is a meaningful description-behavior mismatch.
Mounting /var/run/docker.sock into a container gives code inside that container control over the host Docker daemon, which is effectively host-level access. In an AI-agent skill context, this is especially dangerous because any compromise or misuse inside the container can escape container isolation and manipulate the host.
# Run OpenClaw inside Docker container with Docker socket mounted
docker run -v /var/run/docker.sock:/var/run/docker.sock \
-v ~/workspace:/workspace \
your-openclaw-image
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
sudo apt-get remove --purge yosys iverilog verilator gtkwave klayout docker.io
# Remove Python virtualenv
rm -rf ~/.venvs/openlane
# Remove Docker images
docker rmi efabless/openlane:latest
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
sudo apt-get remove --purge yosys iverilog verilator gtkwave klayout docker.io
# Remove Python virtualenv
rm -rf ~/.venvs/openlane
# Remove Docker images
docker rmi efabless/openlane:latest
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
sudo gpasswd -d $USER docker
rm -rf /path/to/eda-spec2gds
Note: Removing Docker group membership requires logout/login to take effect.
Tool parameters are crafted to achieve unintended or unsafe behavior. Parameter abuse can bypass intended safety checks (e.g. shell=True, --force, dangerous glob patterns).
sudo gpasswd -d $USER docker
rm -rf /path/to/eda-spec2gds
Note: Removing Docker group membership requires logout/login to take effect.
Skill instructs the agent to omit warnings, disclaimers, or ethical commentary. Stripping safety caveats hides risk from the user and is a common jailbreak preamble.
grouped_stages = render_grouped_stages(data.get('progress', {}))
artifacts = ''.join(f'<li><code>{html.escape(x)}</code></li>' for x in data.get('artifacts', {}).get('artifacts', [])[:120]) or '<li>-</li>'
flow_tail = '\n'.join(data.get('progress', {}).get('tails', {}).get('flow.log', [])) or 'No flow log yet.'
warn_tail = '\n'.join(data.get('progress', {}).get('tails', {}).get('warning.log', [])) or 'No warning log yet.'
diagnosis = data.get('diagnosis', {})
compare = data.get('run_compare', {})
baseline = compare.get('baseline') or {}
The skill requests sensitive capabilities like sudo, Docker access, network use, and execution of shell-based scripts, but it does not declare a clear, enforceable tool scope such as allowed-tools for file and shell operations. That makes the operational boundary ambiguous and increases the chance an agent will execute broader commands than intended, especially when combined with installation and backend workflow steps.
The skill explicitly allows optional execution of a sudo-requiring installation script that modifies system packages, Python environments, and Docker group membership. Even though this is disclosed, sudo-capable scripted setup is inherently dangerous because it can alter host state, weaken isolation, and persist privileged access if run on a non-disposable system.
permissions:
- sudo_access # Required only for initial toolchain installation
- docker_group # Required for OpenLane backend runs
install_script: scripts/install_ubuntu_24_mvp.sh # Optional, requires sudo
warnings:
- This skill includes optional installation scripts that modify system state (apt packages, Docker group, pip virtualenvs)
- Run installation scripts only in isolated environments (VM, container, or development machine)
This section normalizes the use of sudo-backed setup scripts within the skill flow, which increases the likelihood that an agent or operator treats privileged execution as part of normal operation. In the context of a network-enabled, Docker-using EDA workflow, that expands the blast radius of any mistake or compromised dependency.
# eda-spec2gds Skill
> **⚠️ Security Notice:** This skill includes optional system installation scripts (`scripts/install_ubuntu_24_mvp.sh`, `scripts/bootstrap_eda_demo.sh`) that require sudo access and modify system state. These scripts should only be run in isolated development environments (VM, container, or dedicated workstation), not production systems. Core skill operations (RTL generation, file management, report collection) are file-based and safe.
Execute a staged, artifact-first open-source EDA flow within the workspace. Prefer deterministic scripts for execution, keeping the agent focused on planning, generation, diagnosis, and iteration.
Pulling efabless/openlane:latest is unsafe because the latest tag is mutable and may resolve to different images over time, including a compromised or unexpected build. In a workflow that may be run with Docker privileges and network access, this creates a serious supply-chain risk and undermines reproducibility.
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
| Dashboard Serving | Local HTTP server on specified port | ✅ Safe |
These operations:
- Only read/write within the skill's project directories
- Do not require network access (except optional dashboard)
- Do not require elevated privileges
- Do not modify system configuration
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
- Only read/write within the skill's project directories
- Do not require network access (except optional dashboard)
- Do not require elevated privileges
- Do not modify system configuration
---
sudo systemctl enable --now docker enables and starts a persistent privileged service. In the context of this skill, that materially changes host state and broadens attack surface beyond simple file-based operations, especially because later guidance also involves Docker socket and group access.
sudo apt-get install yosys iverilog verilator gtkwave klayout docker.io python3-pip python3-venv
# Enables Docker service
sudo systemctl enable --now docker
# Adds user to Docker group (privilege escalation potential)
sudo usermod -aG docker $USER
systemctl enable --now docker introduces persistence by enabling a background service across reboots. While not malware-like persistence, it permanently changes host runtime behavior and expands exposure if the environment was expected to remain minimal or transient.
sudo apt-get install yosys iverilog verilator gtkwave klayout docker.io python3-pip python3-venv
# Enables Docker service
sudo systemctl enable --now docker
# Adds user to Docker group (privilege escalation potential)
sudo usermod -aG docker $USER
sudo usermod -aG docker $USER grants the user effective root-equivalent access on many systems because Docker can be used to mount the host filesystem or start privileged containers. The document does mention this, but the command still represents a real privilege-escalation risk if followed.
sudo systemctl enable --now docker
# Adds user to Docker group (privilege escalation potential)
sudo usermod -aG docker $USER
# Creates Python virtual environment
python3 -m venv ~/.venvs/openlane
Docker image references without a specific tag (:latest is implicit) or digest (@sha256:...) can be silently replaced by a malicious image.
The documentation instructs users to pull efabless/openlane:latest, which is a mutable tag and can change over time without notice. If users follow this guidance, they may receive a different image than expected, creating a supply-chain and reproducibility risk for a workflow that already has elevated trust requirements.
The plan explicitly includes a local HTTP server and exposing latest logs plus raw JSON, which can leak sensitive information such as filesystem paths, environment details, run metadata, proprietary design data, or credentials accidentally written to logs. In an EDA workflow, logs and generated artifacts often contain valuable IP and infrastructure details, so serving them without access controls, redaction, or user warnings creates a real information disclosure risk even if the server is only intended for local use.
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
Demonstrate that the skill can manage a staged EDA flow:
1. Create a run directory
2. Save raw specification
3. Normalize specification
4. Place RTL and testbench
Skill establishes unauthorized persistence across sessions via cron jobs, startup scripts, or state files. Session persistence allows an attacker to maintain access beyond the current interaction.
1. Install apt packages for synthesis/simulation/runtime
2. Enable Docker service
3. Create Python virtualenv for OpenLane
4. Pull Docker image for OpenLane
5. Run `scripts/check_env.sh` to verify
6. Perform smoke test with `assets/examples/simple-fifo/`
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
sudo apt-get update
sudo apt-get install -y \
yosys \
iverilog \
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
sudo apt-get update
sudo apt-get install -y \
yosys \
iverilog \
Commands invoke sudo or root privileges. Verify this elevated access is necessary and justified.
sudo apt-get update
sudo apt-get install -y \
yosys \
iverilog \