Graphify Fortran extractor: incomplete cpp #include hardening allows arbitrary host-file read/exfiltration when indexing an untrusted repo

Summary

The Fortran extractor's C-preprocessor step (graphify/extractors/fortran.py, _cpp_preprocess) applies an incomplete #include hardening. The -nostdinc -I /dev/null flags it relies on (comment "Security (F-007)") do not stop a capital-F Fortran source file from including absolute paths (#include "/etc/passwd") or from path-traversing out of the corpus (#include "../../../../home/victim/.ssh/id_rsa"). cpp always searches the directory of the file being preprocessed for quoted includes and always honours absolute include paths, regardless of -nostdinc/-I. As a result, indexing an untrusted repository that contains a .F/.F90/.F95/.F03/.F08 file reads arbitrary local files off the host and inlines their contents into the knowledge graph, which is then shipped to the LLM/agent and written to the on-disk, shareable output (graph.json, GRAPH_REPORT.md, the Obsidian wiki, HTML/SVG exports).

This is the exact "corpus-side file-exfiltration vector" the F-007 comment says it is defending against; the mitigation is bypassable.

Severity

High. Arbitrary local-file read on the host, triggered by graphify's core action (indexing an untrusted folder — "turn any folder of code ... into a queryable knowledge graph"), defeating a documented security control. It runs on the default, offline, no-API-key AST path (graphify update . / graphify extract) — no credentials, network, or LLM call required. CWE-73 (external control of file name/path) / CWE-22 (path traversal) / CWE-200 (information exposure).

Where

graphify/extractors/fortran.py, _cpp_preprocess():

result = subprocess.run(
    ["cpp", "-w", "-P", "-nostdinc", "-I", "/dev/null", str(path.resolve())],
    capture_output=True, timeout=30,
)

Reached from extract_fortran() (same file) for any suffix in _FORTRAN_CPP_EXTS = {".F", ".F90", ".F95", ".F03", ".F08"}, which are collected by default (detect.CODE_EXTENSIONS) and dispatched by extract._DISPATCH.

Why the mitigation does not work

-nostdinc only removes the standard system include directories, and -I /dev/null only sets the additional search path. Neither affects:

1. Absolute includes — #include "/etc/passwd" is opened directly by cpp.
2. Quoted includes relative to the current file — cpp always searches the directory containing the file being preprocessed for #include "...", so #include "../../../../etc/shadow" traverses out of the corpus.

Proof of concept (verified)

Reproduced against the shipped code with the exact command _cpp_preprocess builds (cpp -w -P -nostdinc -I /dev/null <file>), and against extract_fortran() directly.

Secret on disk (/tmp/secret.txt): TOP_SECRET_...

(a) Absolute include — evil_abs.F90:

program evil
#include "/absolute/path/to/secret.txt"
end program evil

(b) Relative traversal — evil_rel.F90:

program evil
#include "../secret.txt"
end program evil

Both produce preprocessed output containing the full secret file contents ("SECRET IN OUTPUT: True" for both).

Surfacing into the graph — leak.F90 places the include in an identifier position via a free-form continuation so the leaked value becomes a node/edge:

module m
contains
subroutine s
use &
#include "/path/to/token.secret"
end subroutine s
end module m

extract_fortran() on this file emits a node whose label is the entire single-line contents of token.secret and an imports edge to it. That node is written to graph.json / GRAPH_REPORT.md / the wiki and returned by the MCP query_graph / get_node / get_neighbors tools — i.e. exfiltrated into the agent's context and into persisted, shareable artifacts.

Fidelity note: Fortran is case-insensitive, so captured identifiers are lower-cased. Case-insensitive secrets (file paths, hostnames, usernames, hex tokens, connection-string fragments, and the mere existence/partial content of a readable file) leak cleanly; high-entropy case-sensitive blobs (base64/JWT) are partially corrupted by the lower-casing but still confirm readability and structure. Single-token-per-line secret files (API-key files, single-line credential/token files) leak that token in full (lower-cased). Requires cpp/gcc on PATH (near-ubiquitous on the dev/CI hosts graphify targets); it cleanly falls back to raw bytes when absent.

Suggested fix

Mirror the containment approach already used elsewhere in the codebase (_project_xml_is_safe, security.validate_graph_path): before invoking cpp, scan the source for #include directives and refuse or neutralise any include whose target is absolute or resolves outside the corpus/repo root. Concretely, in _cpp_preprocess:

• Parse #include "..." / #include <...> directives.
• Resolve each target against the file's directory; if it is absolute, or its .resolve() escapes the scan root, strip that line (or fall back to path.read_bytes() with no preprocessing).
• Only allow includes that resolve to a path inside the corpus root.

A blunter but safe alternative is to strip all #include lines before preprocessing (macros still expand; cross-file includes lose fidelity but no host files can be read). Running cpp inside a filesystem sandbox (bwrap/unshare with only the corpus mounted) would also close it but is heavier.

Detected by

Aeon vuln-scanner. Semgrep flagged the Fortran subprocess path; manual review of the F-007 mitigation plus an empirical cpp repro confirmed the bypass.