GHSA-vcv2-r9jh-99m5: Agentic-Flow: OS Command Injection in agentic-flow MCP server tools via unsanitized tool-parameter interpolation into execSync

Many tools build a command string like npx --yes agentic-flow --agent "${agent}" --task "${task}" and hand it to execSync(), which uses /bin/sh -c. A task value of x"; touch /tmp/INJECTED; id > /tmp/rce.txt; echo " closes the quote and adds new commands. The advisory's proof of concept creates the marker file and writes id output.

The affected code is wide: 14 tools in src/mcp/standalone-stdio.ts (including agentic_flow_agent, agent_booster_edit_file and agentdb_pattern_store), the FastMCP servers, tools/agent/{execute,list,parallel}.ts, tools/swarm/orchestrate.ts and the depth path of tools/hooks/pretrain.ts. This was a partial-fix gap: an earlier commit, 6a06854, fixed CWE-78 elsewhere but missed the MCP files.

The attacker does not need host access. The AI agent treats these arguments as data, so a web page, file or third-party tool output that steers the agent's tool call can deliver the payload (the vector sets UI:R). The HTTP/SSE transports in http-sse.ts and http-streaming-updated.ts have the same sinks with no authentication and no Origin or Host validation. If they are bound to a reachable interface, a network client can send the payload directly. The result is code execution with the privileges of the MCP server user.

agentic-flow 2.0.13 and earlier. The downstream packages ruflo, claude-flow and @claude-flow/cli were bumped from 3.12.3 to 3.12.4 to pull in the fix, so 3.12.3 and earlier builds that bundle it are likely affected as well. Check:

npm ls agentic-flow
npm ls -g agentic-flow ruflo claude-flow @claude-flow/cli

The MCP server is the exposed part, so any host that runs it is in scope.

Upgrade to agentic-flow 2.0.14.

npm install [email protected]

The fix (pull request 170) rewrites every affected call site to execFileSync(file, argv, { shell: false }). Argument values go straight to execve(2) without shell parsing. For the downstream tools, use [email protected], [email protected] or @claude-flow/[email protected]. npx ...@latest resolves to those. The advisory says no configuration mitigates the flaw without upgrading. If you cannot upgrade, do not bind the HTTP/SSE transports to a reachable interface.

Look at the process tree under the MCP server for sh -c children that run more than the expected npx --yes agentic-flow command, such as touch, curl or id. Search for the PoC marker files /tmp/INJECTED and /tmp/rce.txt. Check MCP tool-call arguments for "; sequences, backticks or $( inside agent, task, name or language. Beam records agent shell commands and MCP tool calls, and its rules flag risky ones, so injected commands and the tool call that carried them can be traced.

Summary

agentic-flow versions <= 2.0.13 MCP server tools interpolated attacker-influenceable tool parameters (e.g. agent, task, name, language, agentdb arguments) directly into shell command strings passed to execSync(). A malicious value reaching any of the affected MCP tools could break out of the surrounding double-quoted argument and execute arbitrary OS commands with the privileges of the user running the MCP server.

This was a partial-fix gap: prior commit 6a06854 (#158) fixed CWE-78 elsewhere in the project but missed the MCP server files entirely.

Impact

Any MCP tool argument that the AI agent treats as data but the implementation interpolates into a shell command string becomes a command-injection vector. In MCP deployments where untrusted content (web pages, files, third-party tool output) is processed by the agent, this is reachable without direct attacker access to the host. The HTTP/SSE transports (http-sse.ts, http-streaming-updated.ts) expose the same sinks without authentication or Origin/Host validation, which may raise the effective severity in any deployment that binds them to a reachable network interface.

Affected components

• src/mcp/standalone-stdio.ts — agentic_flow_agent, agentic_flow_create_agent, agentic_flow_list_all_agents, agentic_flow_agent_info, agentic_flow_check_conflicts, agentic_flow_optimize_model, agentic_flow_list_agents, agent_booster_edit_file, agent_booster_batch_edit, agent_booster_parse_markdown, agentdb_stats, agentdb_pattern_store, agentdb_pattern_search, agentdb_pattern_stats
• src/mcp/fastmcp/servers/claude-flow-sdk.ts
• src/mcp/fastmcp/servers/stdio-full.ts
• src/mcp/fastmcp/servers/http-streaming-updated.ts
• src/mcp/fastmcp/servers/http-sse.ts
• src/mcp/fastmcp/servers/poc-stdio.ts
• src/mcp/fastmcp/tools/agent/{execute,list,parallel}.ts
• src/mcp/fastmcp/tools/swarm/orchestrate.ts
• src/mcp/fastmcp/tools/hooks/pretrain.ts (depth path only)

Proof of Concept

// Pre-fix (standalone-stdio.ts, agentic_flow_agent)
let cmd = `npx --yes agentic-flow --agent "${agent}" --task "${task}"`;
const result = execSync(cmd, { encoding: 'utf-8', ... });

Invoking the MCP tool with:

{
  "agent": "coder",
  "task": "x\"; touch /tmp/INJECTED; id > /tmp/rce.txt; echo \""
}

produces, after interpolation:

npx --yes agentic-flow --agent "coder" --task "x"; touch /tmp/INJECTED; id > /tmp/rce.txt; echo ""

When execSync hands that to /bin/sh -c, the shell parses three commands: the truncated npx, then touch /tmp/INJECTED, then id > /tmp/rce.txt; echo "". The marker file /tmp/INJECTED is created and the user's id output is written to /tmp/rce.txt.

Patches

Fixed in `[email protected]` — every affected call site rewritten to use execFileSync(file, argv, { shell: false }) so attacker-controlled argv elements are passed straight to execve(2) without shell parsing.

Fix PR: ruvnet/agentic-flow#170 (merged at 0c2ec96)

A regression test (tests/security/cwe-78-mcp-execsync.test.ts) was added that statically scans every src/mcp/**/*.ts file and fails the build if any new execSync() call is reintroduced outside of a documented exemption, plus a behavioural smoke check that the canonical PoC payload remains inert when passed as an argv element to execFileSync.

Workarounds

Upgrade to agentic-flow >= 2.0.14. There is no in-product configuration that mitigates this without upgrading.

Downstream pin

The ruflo / claude-flow / @claude-flow/cli packages bumped from 3.12.3 → 3.12.4 to pull the patched agentic-flow:

• [email protected]
• [email protected]
• @claude-flow/[email protected]

End users running any of npx ruflo@latest, npx claude-flow@latest, or npx @claude-flow/cli@latest are pinned to the fixed version.

Credit

Reported by hackchang via a well-scoped red-team report package (npm_agentic-flow_report_package_20260618_163017.zip) that included a sink inventory, a minimized PoC payload, and a clear explanation of why this was a partial-fix gap rather than intended behaviour. The sink inventory directly drove the single-grep pass that closed every reachable call site; the PoC payload became the behavioural smoke test that proves the canonical attack stays inert as an argv element.