Back to skill

Security audit

RDK YOLO Toolkit

Security checks for vulnerabilities and agentic risk

Overview

This skill is mostly aligned with RDK X5 YOLO deployment, but it gives agents broad local and SSH command authority, starts detached camera/inference jobs, and includes unsafe installation and device-permission guidance.

Install only if you are comfortable letting an agent run local and SSH commands on training machines and RDK X5 boards. Prefer the command-list/manual mode, review every ssh/scp/nohup/tmux command before execution, avoid root unless required, replace chmod 666 with a dedicated group or udev rule, and harden the installer by pinning versions and verifying hashes before running it.

Vulnerability Patterns
  • Remote Payload Retrieval and ExecutionFetches external code whose behavior can change after review
  • Unauthorized Access and Privilege EscalationObtains permissions beyond the task's legitimate needs
  • Insecure DependenciesIntroduces malicious components through unsafe dependency sources
  • Skill Instruction HijackingAlters the agent's session goals or safety constraints when the skill loads
  • Agent Memory PoisoningWrites attacker-controlled rules into memory that affect later sessions
Findings (3)

T03 · Remote Payload Retrieval and Execution

Error
Location
scripts/install_train_env.sh:10
Finding

Unverified Mutable Miniconda Installer Is Downloaded and Executed

Content
View full analysis

Vulnerability Details

File Location: scripts/install_train_env.sh, lines 10–16
Vulnerability Type: Remote payload retrieval and execution without integrity verification
Risk Level: High

bash
if [ ! -d ~/miniconda3 ]; then
    echo "--- Install Miniconda ---"
    cd ~
    if [ ! -f Miniconda3-latest-Linux-x86_64.sh ]; then
        wget -q https://repo.anaconda.com/miniconda/Miniconda3-latest-Linux-x86_64.sh
    fi
    bash Miniconda3-latest-Linux-x86_64.sh -b -p ~/miniconda3
fi

Technical Analysis

The installation script downloads a mutable latest artifact and executes it without validating a cryptographic checksum or signature. HTTPS protects the connection in ordinary circumstances but does not establish that the downloaded artifact is the exact version reviewed by the project.

The script also trusts an existing file named ~/Miniconda3-latest-Linux-x86_64.sh. If that predictable file is already present, the download is skipped and the existing file is executed. Consequently, a local attacker or another compromised process with write access to the account’s home directory can replace the installer before this script runs.

Because the external artifact can change independently of this repository, the effective executable payload cannot be determined solely by auditing the project.

Attack Path

  1. An attacker compromises the upstream distribution channel, the retrieved artifact, or a process with write access to the target user’s home directory.
  2. In the local planting scenario, the attacker creates a malicious file at ~/Miniconda3-latest-Linux-x86_64.sh.
  3. The user or Agent runs scripts/install_train_env.sh, either locally or through the documented SSH workflow.
  4. The script detects the existing predictable filename and skips downloading a clean copy, or downloads a compromised mutable artifact.
  5. The script invokes the file with bash, executing the attacker’s command ...[truncated 508 chars]
Remediation
View remediation

Remediation Suggestions

  1. Pin a specific Miniconda release rather than using the mutable latest filename.
  2. Store the expected vendor-published SHA-256 or stronger digest in the repository and verify it before execution.
  3. Download into a private temporary directory created with mktemp -d, rather than into the user’s home directory.
  4. Use a fixed output filename and fail closed if downloading or validation fails.
  5. Do not execute a pre-existing installer unless it passes the same integrity check.
  6. Remove the temporary installer after successful installation.
  7. Avoid running the setup script as root.

Example hardening pattern:

bash
version="py310_25.1.1-2"
expected_sha256="VENDOR_PUBLISHED_SHA256"
tmp_dir="$(mktemp -d)"
trap 'rm -rf "$tmp_dir"' EXIT
installer="$tmp_dir/miniconda.sh"

wget --https-only \
  "https://repo.anaconda.com/miniconda/Miniconda3-${version}-Linux-x86_64.sh" \
  -O "$installer"

printf '%s  %s\n' "$expected_sha256" "$installer" | sha256sum --check -
bash "$installer" -b -p "$HOME/miniconda3"

T08 · Insecure Dependencies

Warning
Location
scripts/install_train_env.sh:33
Finding

Unpinned Python Dependencies Permit Unreviewed Package Changes

Content
View full analysis

Vulnerability Details

File Location: scripts/install_train_env.sh, lines 33–39
Vulnerability Type: Insecure and non-reproducible third-party dependency installation
Risk Level: Medium

bash
PIP_OPTS="-i https://pypi.tuna.tsinghua.edu.cn/simple --timeout 60 --retries 10"
echo "--- Install PyTorch (CUDA 12.1) ---"
pip install $PIP_OPTS torch torchvision --index-url https://download.pytorch.org/whl/cu121
echo "--- Install ultralytics + rdkx5-yolo-mapper + onnx ---"
pip install $PIP_OPTS "ultralytics>=8.3.0" rdkx5-yolo-mapper onnx onnxsim
echo "--- Install missing setuptools dependency for rdkx5-yolo-mapper ---"
pip install $PIP_OPTS setuptools

Technical Analysis

The installer does not pin exact versions for torch, torchvision, rdkx5-yolo-mapper, onnx, onnxsim, or setuptools. The ultralytics requirement has only a lower bound and therefore permits any future version greater than or equal to 8.3.0.

No package hashes are verified. The resulting environment depends on whichever distributions and transitive dependencies are served by the configured indexes at installation time. Python package installation can execute build backend logic and other package-supplied code, so a compromised package or index can lead to code execution during installation.

The use of both a third-party mirror and the PyTorch index increases the set of external supply-chain systems trusted by the script. The lack of a lock file also prevents reliable reproduction and review of the resolved dependency graph.

Attack Path

  1. An upstream project account, package release, package index, mirror, or transitive dependency is compromised.
  2. A malicious or vulnerable release satisfying the broad requirement is published.
  3. The user or Agent runs the environment installation script.
  4. pip resolves the unpinned requirement to the altered release and downloads it.
  5. Package installation or later import exec ...[truncated 496 chars]
Remediation
View remediation

Remediation Suggestions

  1. Pin every direct dependency to an exact reviewed version.
  2. Generate a lock file that also pins all transitive dependencies.
  3. Record and enforce distribution hashes with pip install --require-hashes.
  4. Explicitly separate packages that must come from the PyTorch index from packages obtained through the primary index.
  5. Prefer an organization-controlled package proxy that validates and retains approved artifacts.
  6. Regularly scan the lock file with dependency and vulnerability analysis tools.
  7. Test dependency upgrades in isolation and update pins only after review.

A hardened installation should follow this form:

bash
python -m pip install \
  --require-hashes \
  --only-binary=:all: \
  --no-deps \
  -r requirements.lock

The lock file should contain exact versions and hashes for every resolved package. Platform-specific packages that cannot use --only-binary=:all: should be separately reviewed and built in a controlled environment.

T05 · Unauthorized Access and Privilege Escalation

Warning
Location
SKILL.md:828
Finding

Recommended World-Writable Permissions Expose Shared-Memory Devices

Content
View full analysis

Vulnerability Details

File Location: SKILL.md, line 828
Vulnerability Type: Excessive device permissions and violation of least privilege
Risk Level: Medium

bash
sudo chmod 666 /dev/hbmem*

Technical Analysis

Permission mode 0666 grants read and write access to every local user for every device node matching /dev/hbmem*. Device nodes are security-sensitive kernel interfaces and should not be made globally writable merely to permit one ROS2 process to use them.

The documentation mentions group membership as an alternative, but it still presents world-writable permissions as a valid remediation. Until the permissions are reset or the device nodes are recreated, unrelated local processes can attempt to interact with the shared-memory devices.

Attack Path

  1. An administrator follows the documented recommendation and runs sudo chmod 666 /dev/hbmem*.
  2. An unprivileged local user or compromised low-privilege process discovers the now-readable and writable device nodes.
  3. That process opens the device interface without belonging to an authorized hardware-access group.
  4. Subject to driver behavior, the process reads, corrupts, injects, or disrupts data used by camera and inference workloads.

Impact Assessment

The command does not directly grant full root privileges, but it breaks the intended access-control boundary around the shared-memory hardware interface. Potential consequences include exposure of camera or inference data, corruption of shared buffers, interference with ROS2 processing, denial of service, and exploitation of any reachable device-driver vulnerabilities. The scope includes all local users and processes while the permissive mode remains active.

Remediation
View remediation

Remediation Suggestions

  1. Remove the recommendation to use mode 0666.
  2. Create a dedicated group, such as hbmem, for processes that require access.
  3. Add only the dedicated ROS2 service account or explicitly authorized users to that group.
  4. Configure a narrowly scoped udev rule that assigns the correct group and mode 0660 when the device nodes are created.
  5. Restrict access to only the required device nodes rather than using an unnecessarily broad wildcard.
  6. Restart the affected session or service after group membership changes instead of weakening permissions globally.
  7. Review the vendor’s recommended permissions and capabilities before deploying a custom udev rule.

Example policy:

text
KERNEL=="hbmem*", GROUP="hbmem", MODE="0660"

The exact match criteria should be verified against the device metadata using udevadm info before deployment.

Vulnerability Patterns
  • Privilege EscalationExcessive Permissions, Sudo/Root Execution, Credential Access
  • Tool MisuseTool Parameter Abuse, Chaining Abuse, Unsafe Defaults
  • Rogue AgentSelf-Modification, Session Persistence
  • Trigger AbuseOverly Broad Trigger, Shadow Command Trigger, Keyword Baiting Trigger
  • YARA SignaturesMalware Match, Webshell Match, Cryptominer Match
Findings (26)

Tp4

High
Category
MCP Tool Poisoning
Confidence
97% confidence
Finding

声明描述的是一个覆盖训练、模型导出、量化部署和 ROS2 实时检测的端到端能力集合;而提供的代码片段只是部署前的环境检查工具。它读取 /etc/version,检查 /opt/tros 下文件、package.xml 版本、Python 模块导入情况以及 /dev/bpu* 设备节点,属于辅助性验证脚本。虽然这些检查与 RDK X5 部署场景相关,但该代码的主要目的与声明的核心功能明显不同,且关键宣称能力均未在代码中体现,因此构成明显不匹配。

Content

No source excerpt is available for this finding.

Tool Parameter Abuse

High
Category
Tool Misuse
Confidence
99% confidence
Finding

The recommended chmod 666 is a classic unsafe parameter pattern because it applies maximally permissive access to device files. In this context, the skill could drive an agent to suggest or execute an overly broad permission change on specialized hardware interfaces, enabling unintended local access and weakening host security significantly.

Content

Scanner excerpt · SKILL.md (reported line 828)May include surrounding context.

pkill -f "dnn_node_example/example" # dnn_node

text

⚠️ **`/dev/hbmem*` 设备权限**:默认 root 才能写 hbmem。非 root 用户需 `sudo chmod 666 /dev/hbmem*` 或加入对应 group。

⚠️ **topic 名对齐总览**:
| 输入源节点 | hbmem 默认 topic | 需改名? |

YARA rule 'backdoor_persistence': Backdoor persistence with malicious payloads (shell commands, SSH key injection, hidden root users) [malware]

High
Category
YARA Match
Confidence
75% confidence
Finding

YARA rule matched a known malware signature (reverse shell, backdoor, ransomware, C2 framework, or info stealer).

Content

Scanner excerpt · scripts/install_train_env.sh (reported line 52)May include surrounding context.

sh
r/share/fonts/truetype/dejavu/DejaVuSans.ttf ~/.config/Ultralytics/Arial.ttf 2>/dev/null || \
    echo "   ⚠️ DejaVu 字体没装,需手动 apt install fonts-dejavu-core"

# 5) WSL2 GPU 检测
if [ -f /usr/lib/wsl/lib/nvidia-smi ] && ! which nvidia-smi >/dev/null 2>&1; then
    echo "--- WSL2 检测到 GPU 但 nvidia-smi 不在 PATH ---"
    grep -q '/usr/lib/wsl/lib' ~/.bashrc || \
        echo 'export PATH=/usr/lib/wsl/lib:$PATH' >> ~/.bashrc
    export PATH=/usr/lib/wsl/lib:$PATH
fi

# 6) 验证(任何一项失败请回到对应步骤修复)
echo ""
echo "=== 验证 ==="
python -c "import torch; print('torch', torch.__version__, 'cuda', torch.cuda.is_available(), torch.cuda.get_device_name(0) if torch.cuda.is_available() else '(no GPU)')"
python -c "import pkg_resources; print('setuptools OK')"
python -c "import ultralytics; print('ultralytics', ultralytics.__version__)"
hb_mapper --version
ls ~/.config/Ultralytics/Arial.ttf >/dev/null && echo "Arial.ttf 占位 ✅"

echo

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
93% confidence
Finding

The skill description is written as a direct user-facing description entirely in Chinese, which implies a fixed interaction language for the skill. Under the policy, language constraints should offer user choice or clearly document a justified locale limitation; this README does neither.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
79% confidence
Finding

The README explicitly promotes remote agent execution over SSH with tmux/nohup, which enables persistent unattended command execution on user infrastructure. In the context of an agent skill that can run Bash remotely, this increases the risk of long-lived processes continuing after user oversight ends, making accidental misuse, resource abuse, or harder-to-detect malicious actions more dangerous.

Content

Scanner excerpt · README.md (reported line 24)May include surrounding context.

md
Agent 会**按节点对话式追问**接入方式(不需要你一次性填表),并按 4 档模式执行:
- **本机直跑** — 你在 GPU 机器上调用,Agent 用 Bash 直接跑
- **SSH 远程** — 你给 ssh,Agent 远程 + tmux/nohup
- **输出指导** — 你说"只要命令",Agent 给你清单 + Checkpoint
- **板端 SSH** — 给 RDK X5 ssh,自动连上去推理 + ROS2

Undeclared Tool Scope

Medium
Category
MCP Least Privilege
Confidence
97% confidence
Finding

The skill instructs the agent to run shell commands locally and over SSH, but it declares no explicit tool scope or allowed-tools boundary. That creates an authorization gap where a broad-capability agent may execute high-impact commands without a least-privilege contract, increasing the chance of unintended command execution on user systems.

Content

No source excerpt is available for this finding.

Natural-Language Policy Violations

Medium
Category
Not specified by scanner
Confidence
97% confidence
Finding

The top-level description is written entirely in Chinese and the operational instructions throughout the file are also Chinese, with no indication that the user can choose another language. This creates a language-policy concern because the skill effectively imposes a locale/language preference without documented opt-in or justification.

Content

No source excerpt is available for this finding.

Vague Triggers

Medium
Category
Not specified by scanner
Confidence
89% confidence
Finding

The trigger condition is very broad and activates on common platform terms like RDK X5, BPU, hb_mapper, or ROS-related keywords. Broad matching can cause the skill to engage in unrelated conversations and escalate into offering shell/SSH actions where the user did not intend operational changes, increasing the risk of accidental execution or over-collection of connection details.

Content

No source excerpt is available for this finding.

Session Persistence

Medium
Category
Rogue Agent
Confidence
85% confidence
Finding

The skill explicitly recommends tmux or nohup for remote execution durability over SSH. In a skill that can run shell commands remotely, backgrounding processes without strong consent and visibility can leave long-running jobs active after the session ends, reducing user control and complicating incident response.

Content

Scanner excerpt · SKILL.md (reported line 89)May include surrounding context.

md
| 档位 | 触发条件 | Agent 行为 |
|---|---|---|
| **本机直跑** | 用户在 GPU 机器/x86 Linux 上调用 Agent | 直接用 `Bash` 工具跑 `yolo train` / `mapper.py` |
| **SSH 远程** | 用户给了 host/user/port | 用 `ssh -p PORT user@host 'bash -s' < script.sh` 远程跑,tmux 或 nohup 防断开 |
| **输出指导** | 用户明说"只要命令" | 输出完整命令清单 + 校验点,要求用户回报关键指标(mAP、bin 大小、forward 时间) |
| **板端 SSH** | 用户给了 RDK X5 SSH | ssh + scp 全自动;板端默认无 tmux,用 `nohup ... &` 兜底 |

Session Persistence

Medium
Category
Rogue Agent
Confidence
85% confidence
Finding

The SSH safety section normalizes detached execution with nohup as a standard operating mode. Persistent remote jobs can continue consuming compute, network, cameras, or BPU resources after disconnect, and on shared systems this can create operational and security issues if ownership and termination are unclear.

Content

Scanner excerpt · SKILL.md (reported line 91)May include surrounding context.

md
| **本机直跑** | 用户在 GPU 机器/x86 Linux 上调用 Agent | 直接用 `Bash` 工具跑 `yolo train` / `mapper.py` |
| **SSH 远程** | 用户给了 host/user/port | 用 `ssh -p PORT user@host 'bash -s' < script.sh` 远程跑,tmux 或 nohup 防断开 |
| **输出指导** | 用户明说"只要命令" | 输出完整命令清单 + 校验点,要求用户回报关键指标(mAP、bin 大小、forward 时间) |
| **板端 SSH** | 用户给了 RDK X5 SSH | ssh + scp 全自动;板端默认无 tmux,用 `nohup ... &` 兜底 |

### SSH 安全约定
- **不索要明文密码**,假设用户配好 SSH key

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · SKILL.md (reported line 496)May include surrounding context.

md
# Step 1: conda 环境 + ToS(conda 26+ 必须)
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
conda create -n rdk_env python=3.10 -y && conda activate rdk_env

# Step 2: 国内必须换源装 PyTorch(pypi 直连必超时)
PIP_OPTS="-i https://pypi.tuna.tsinghua.edu.cn/simple --timeout 60 --retries 10"

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · SKILL.md (reported line 496)May include surrounding context.

md
# Step 1: conda 环境 + ToS(conda 26+ 必须)
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/main
conda tos accept --override-channels --channel https://repo.anaconda.com/pkgs/r
conda create -n rdk_env python=3.10 -y && conda activate rdk_env

# Step 2: 国内必须换源装 PyTorch(pypi 直连必超时)
PIP_OPTS="-i https://pypi.tuna.tsinghua.edu.cn/simple --timeout 60 --retries 10"

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

This command launches a USB camera ROS2 node under nohup in the background. In context, that means a sensor process can continue capturing and publishing image data after the initiating session ends, which raises privacy and operational control concerns beyond ordinary job persistence.

Content

Scanner excerpt · SKILL.md (reported line 733)May include surrounding context.

bash
# 后台 1: 启动 USB 摄像头节点(默认发到 /hbmem_img,与 dnn_node 默认订阅完美对齐)
nohup ros2 launch hobot_usb_cam hobot_usb_cam.launch.py \
  usb_video_device:=/dev/video8 \
  usb_image_width:=640 usb_image_height:=480 \
  usb_pixel_format:=mjpeg \

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

This starts the dnn_node pipeline as a detached background process. Persistent inference services can continue reading shared-memory topics and processing sensor data after user disconnect, increasing the risk of unnoticed resource use or unintended ongoing analytics.

Content

Scanner excerpt · SKILL.md (reported line 741)May include surrounding context.

md
usb_zero_copy:=True > usb_cam.log 2>&1 &

# 后台 2: dnn_node(无需任何 topic remap,默认订阅 /hbmem_img)
nohup bash -c "cd /home/root/hobot_ws && \
  ros2 run dnn_node_example example --ros-args \
    -p feed_type:=1 -p is_shared_mem_sub:=1 \
    -p config_file:=config/custom_workconfig.json" > dnn.log 2>&1 &

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

The MIPI camera launch under nohup creates the same persistence risk as the USB camera case, but with a board-attached sensor path. Detached sensor publishing can remain active and continuously expose camera-derived data or consume platform resources without an active session.

Content

Scanner excerpt · SKILL.md (reported line 760)May include surrounding context.

bash
# 板端常见 sensor: F37 / GC4663 / IMX415 / OV5647
nohup ros2 launch mipi_cam mipi_cam_640x480_nv12_hbmem.launch.py \
  mipi_video_device:=F37 > mipi_cam.log 2>&1 &
# 或 1920x1080: ros2 launch mipi_cam mipi_cam_1920x1080_nv12_hbmem.launch.py

Session Persistence

Medium
Category
Rogue Agent
Confidence
87% confidence
Finding

This detached dnn_node invocation persists inference after the initiating shell exits. In a hardware-deployment skill with SSH automation, that persistence materially increases the blast radius of accidental or premature command execution because the task keeps running independently.

Content

Scanner excerpt · SKILL.md (reported line 765)May include surrounding context.

md
# 或 1920x1080: ros2 launch mipi_cam mipi_cam_1920x1080_nv12_hbmem.launch.py

# dnn_node(同 A1,无需改话题)
nohup bash -c "cd /home/root/hobot_ws && \
  ros2 run dnn_node_example example --ros-args \
    -p feed_type:=1 -p is_shared_mem_sub:=1 \
    -p config_file:=config/custom_workconfig.json" > dnn.log 2>&1 &

Session Persistence

Medium
Category
Rogue Agent
Confidence
83% confidence
Finding

The image publisher is launched as a background process with nohup. Even though this is a test source rather than a live camera, it still creates unattended data publication and process persistence that may surprise users and consume shared ROS resources.

Content

Scanner excerpt · SKILL.md (reported line 783)May include surrounding context.

bash
# 后台 1:图像源 → /hbmem_img(10 fps 循环)⚠️ 必须显式 publish_message_topic_name:=/hbmem_img
nohup ros2 launch hobot_image_publisher hobot_image_publisher.launch.py \
  publish_image_source:=/home/root/hobot_ws/config/test.jpg \
  publish_image_format:=jpg \
  publish_source_image_w:=640 publish_source_image_h:=480 \

Session Persistence

Medium
Category
Rogue Agent
Confidence
86% confidence
Finding

This background dnn_node command creates a detached inference consumer paired with the test publisher. As with the other nohup patterns, persistence lowers transparency and can leave autonomous processing active beyond the intended troubleshooting session.

Content

Scanner excerpt · SKILL.md (reported line 792)May include surrounding context.

md
publish_is_shared_mem:=True > pub.log 2>&1 &

# 后台 2:dnn_node(同 A1)
nohup bash -c "cd /home/root/hobot_ws && \
  ros2 run dnn_node_example example --ros-args \
    -p feed_type:=1 -p is_shared_mem_sub:=1 \
    -p config_file:=config/custom_workconfig.json" > dnn.log 2>&1 &

Session Persistence

Medium
Category
Rogue Agent
Confidence
82% confidence
Finding

The document broadly recommends nohup for all background nodes as a normal pattern. This institutionalizes persistent execution across multiple services and makes it more likely users will leave remote workloads and sensor-related processes running unintentionally.

Content

Scanner excerpt · SKILL.md (reported line 820)May include surrounding context.

通用注意事项(所有模式)

⚠️ 板端默认没 tmux,用 nohup ... > log 2>&1 & 跑后台节点;清理用:

bash
pkill -f hobot_usb_cam                # USB cam
pkill -f mipi_cam                     # MIPI cam

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
99% confidence
Finding

Using chmod 666 on hardware-related device files makes them readable and writable by every local user and process. In the context of ROS2 and shared-memory image transport, that can allow unauthorized consumers or producers to interact with sensitive data paths and disrupt inference pipelines.

Content

Scanner excerpt · SKILL.md (reported line 828)May include surrounding context.

pkill -f "dnn_node_example/example" # dnn_node

text

⚠️ **`/dev/hbmem*` 设备权限**:默认 root 才能写 hbmem。非 root 用户需 `sudo chmod 666 /dev/hbmem*` 或加入对应 group。

⚠️ **topic 名对齐总览**:
| 输入源节点 | hbmem 默认 topic | 需改名? |

Sudo/Root Execution

Medium
Category
Privilege Escalation
Confidence
99% confidence
Finding

Using chmod 666 on hardware-related device files makes them readable and writable by every local user and process. In the context of ROS2 and shared-memory image transport, that can allow unauthorized consumers or producers to interact with sensitive data paths and disrupt inference pipelines.

Content

Scanner excerpt · SKILL.md (reported line 828)May include surrounding context.

pkill -f "dnn_node_example/example" # dnn_node

text

⚠️ **`/dev/hbmem*` 设备权限**:默认 root 才能写 hbmem。非 root 用户需 `sudo chmod 666 /dev/hbmem*` 或加入对应 group。

⚠️ **topic 名对齐总览**:
| 输入源节点 | hbmem 默认 topic | 需改名? |

Session Persistence

Medium
Category
Rogue Agent
Confidence
65% confidence
Finding

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.

Content

Scanner excerpt · SKILL.md (reported line 883)May include surrounding context.

md
echo "7) tmux(可选)"
which tmux >/dev/null 2>&1 \
    && echo "   ✅ tmux 已装" \
    || echo "   ⚠️  tmux 未装(不重要,Agent 会用 nohup 兜底)"

echo ""
echo "=== 自检结束 ==="

Session Persistence

Medium
Category
Rogue Agent
Confidence
65% confidence
Finding

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.

Content

Scanner excerpt · SKILL.md (reported line 884)May include surrounding context.

md
echo "7) tmux(可选)"
which tmux >/dev/null 2>&1 \
    && echo "   ✅ tmux 已装" \
    || echo "   ⚠️  tmux 未装(不重要,Agent 会用 nohup 兜底)"

echo ""
echo "=== 自检结束 ==="

Session Persistence

Medium
Category
Rogue Agent
Confidence
65% confidence
Finding

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.

Content

Scanner excerpt · scripts/check_rdkx5_env.sh (reported line 63)May include surrounding context.

sh
echo "7) tmux(可选)"
which tmux >/dev/null 2>&1 \
    && echo "   ✅ tmux 已装" \
    || echo "   ⚠️  tmux 未装(不重要,Agent 会用 nohup 兜底)"

echo ""
echo "=== 自检结束 ==="

Session Persistence

Medium
Category
Rogue Agent
Confidence
60% confidence
Finding

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.

Content

Scanner excerpt · scripts/install_train_env.sh (reported line 28)May include surrounding context.

sh
# 2) 创建 rdk_env
if ! conda env list | grep -q '^rdk_env '; then
    echo "--- 创建 rdk_env (python 3.10) ---"
    conda create -n rdk_env python=3.10 -y
fi
conda activate rdk_env