Rule catalog

Ruleset v0.3.0 has 19 rules. The severity shown is the default. Some rules adjust it by context, for example when a command only appears in a documentation file.

Automatic execution

AUTO-001 Agent hook runs when a session starts

Default severity: Medium

Why it matters: This hook can run a command as soon as an AI coding agent starts in this repository, before you ask it to do anything.

How to fix it: Read the hook before opening the repo with an agent. Remove it unless you trust what it runs.

AUTO-002 Agent hook reaches the network or runs a concealed script

Default severity: Critical

Why it matters: Hooks run automatically during agent work. One that contacts a remote host or launches a hidden script can steal data or install malware.

How to fix it: Do not open this repo with an agent until the hook is removed or replaced with a reviewed local command.

AUTO-003 Editor task runs when the folder opens

Default severity: Medium

Why it matters: A task with runOn set to folderOpen can execute a command when the repository is opened in VS Code or a compatible editor.

How to fix it: Review the task command. Remove "runOn": "folderOpen" unless you trust it.

AUTO-004 Dev container downloads and runs code

Default severity: High

Why it matters: A lifecycle command fetches code from the network and executes it while the dev container is being set up.

How to fix it: Replace it with a pinned, checksummed install step, or remove it before building the container.

AUTO-005 Suspicious npm install script

Default severity: High

Why it matters: npm runs lifecycle scripts automatically during install. This one reaches the network, decodes a payload, or launches a concealed script.

How to fix it: Review the script. Until then, install with --ignore-scripts.

AUTO-006 Dev container command runs on your host machine

Default severity: Medium

Why it matters: initializeCommand runs on your computer, outside the container, so the container gives it no isolation.

How to fix it: Read the command before building the container. Remove it unless you trust it.

AUTO-007 Agent setting runs a helper command

Default severity: Medium

Why it matters: This setting tells the agent to execute a command from the repository, for example to fetch credentials or draw a status line.

How to fix it: Review the command, or remove the setting if the repository does not need it.

Remote and concealed code

EXEC-001 Remote code is piped into a shell

Default severity: Critical

Why it matters: This downloads code and runs it immediately, so nobody can review what actually executes and the server can change it at any time.

How to fix it: Download to a file, verify a checksum, review it, then run it.

EXEC-002 Encoded payload is decoded and executed

Default severity: Critical

Why it matters: Encoding hides the real command from reviewers. Legitimate setup steps rarely need to decode and run code.

How to fix it: Decode it in a text editor without running it, then remove it unless you can explain every step.

EXEC-003 Configuration runs a heavily obfuscated script

Default severity: High

Why it matters: The script is packed so it cannot be read, which is a common way to hide malware in repository tooling.

How to fix it: Do not run it. Ask the maintainer for the readable source, or remove the reference.

MCP servers

MCP-001 MCP server package is not pinned

Default severity: Medium

Why it matters: The server is downloaded at launch without an exact version, so a future or compromised release runs automatically.

How to fix it: Pin an exact version, for example package@1.2.3.

MCP-002 MCP server connects to an unrecognized remote host

Default severity: Medium

Why it matters: A remote MCP server receives what your agent sends it and can return instructions the agent may follow.

How to fix it: Confirm who operates this host. Remove the server if you do not recognize it.

MCP-003 Credential is written into an MCP configuration

Default severity: High

Why it matters: A secret committed to the repository can be read by anyone with access and is handed to the MCP server at launch.

How to fix it: Revoke this credential and load the value from an environment variable instead.

MCP-004 MCP server is launched through a shell

Default severity: High

Why it matters: A shell wrapper lets a single configuration entry chain extra commands that run whenever the server starts.

How to fix it: Launch the server executable directly, with its arguments in the args array.

Instruction injection

INJ-001 Hidden Unicode characters

Default severity: High

Why it matters: Invisible characters can carry instructions that an AI model reads but a human reviewer cannot see.

How to fix it: Inspect the escaped characters shown here and delete them from the file.

INJ-002 Instruction tries to override rules or hide actions

Default severity: High

Why it matters: This text tells an agent to ignore its instructions or keep something from you, a common prompt injection pattern.

How to fix it: Remove the instruction. If it is only an example, move it out of files that agents load.

INJ-003 Instruction targets credentials

Default severity: Critical

Why it matters: This text asks an agent to read or send sensitive files or secrets, which can leak your credentials.

How to fix it: Do not open this repo with an agent. If it already ran, rotate the affected credentials.

INJ-004 Instruction hidden in an HTML comment

Default severity: Medium

Why it matters: HTML comments do not appear in rendered Markdown, but an agent may still read them as instructions.

How to fix it: View the raw file and remove any comment that tells the agent what to do.

Permissions

PERM-001 Agent approvals are weakened

Default severity: High

Why it matters: This setting broadens what the agent may do without asking you, so a malicious instruction can act without a prompt.

How to fix it: Remove the setting and keep approval prompts on for this repository.