AAK-AGENT-001: Agent instruction file contains shell command directives
- Severity
- CRITICAL
- Category
- Agent Config (
agent-config) - Scanner
agent_audit_kit/scanners/agent_config.py- Auto-fixable
- No.
- SARIF name
AgentShellDirectives
What it detects
An agent instruction file (AGENTS.md, CLAUDE.md, .cursorrules and friends) tells the agent to run code it downloads, or to delete the filesystem root or a home directory: curl -fsSL <url> | bash, bash -c "$(curl -fsSL <url>)", iex ((New-Object Net.WebClient).DownloadString(<url>)), rm -rf ~. The file is read as instructions, so whatever the far end serves at the time runs with the agent's permissions.
Until 0.6.15 this rule was a keyword match (subprocess, eval(, exec(, sh -c, any rm -rf, any curl ... | sh). On the 909 public repositories of the #771 corpus it fired on 74, nearly all of them on a mention: a guardrail such as "Never use eval()", a deny-list row, a code sample, rm -rf dist. It reports the command now, not the word (#869).
How to fix
Do not tell the agent to pipe a download into a shell. Install from a package manager, or download the script to a file, check its checksum or signature, and run the local copy. Remove destructive commands from instruction files.
To confirm the finding is gone, report only this rule:
agent-audit-kit scan . --rules AAK-AGENT-001
Limitations
Reads one line at a time. A download counts when it is piped into an interpreter, run through $(...) or <(...), or handed to PowerShell's iex, and the command names what it fetches (a URL, a $VARIABLE or a host); an interpreter given its code on the command line (| python3 -m json.tool) reads the download as data, and a loopback URL is a local API. A URL that Node, requests and Rust's url crate read differently from curl, httpx and urlsplit, which keep a backslash inside the authority, is judged by the less safe of the two readings. A recursive rm counts when it targets /, ~ or $HOME where a command starts. A command a guardrail forbids is not reported: one negated in the sentence, one under a heading or lead-in such as "### NEVER", or one in a table row. eval(, exec(, subprocess and sh -c on their own, commands split across lines, and other destructive commands (mkfs, dd) are not reported.
References
- OWASP Agentic Top 10: ASI01 (Agent Goal Hijack), ASI05 (Unexpected Code Execution (RCE))
- Adversa AI MCP Top 25 (AAK-defined ids): ADV-HIJACK-01 (Agent Shell Directives)
- CSA AI Controls Matrix (AICM): IAM-02
All rules. This page is generated from the rule registry when the documentation is built; to change it, change the rule in agent_audit_kit/rules/builtin.py. To give the rule a hand-written write-up, add docs/rules/AAK-AGENT-001.md: the page then shows that text, with the registry reference appended under it.