Terminal Injection via Detection Output (Issue #498)#

Magic-rule description output built from file content is not escaped for terminal control bytes. This is a distinct and open vulnerability compared to the TTY-gated escaping that exists for symlink-target rendering.

The Vulnerability#

When a magic rule like 0 string x STR=[%s] reads file content into a description, the raw bytes — including OSC sequences, SGR escape codes, and newlines — reach stdout unescaped. A planted file processed by rmagic * can:

• Spoof the terminal title via OSC sequences (e.g., ESC ] 0 ; TITLE BEL)
• Execute OSC 52 clipboard commands
• Produce multi-line output that overflows the expected single-line result

Confirmed in issue #498: a file containing ZZZZ\x1b]0;PWN\x07X\nSECOND\n causes rmagic to emit two lines including the raw ESC sequence, while GNU file stops at the first newline 1.

Root Cause: Conflated Scan and Render Bounds#

rmagic uses EvaluationConfig::max_string_length (default 8192) for both the scan-window cap on read_string/read_pattern_match and the render output bound. GNU file keeps these separate: it caps the rendered field at 127 bytes and stops at the first newline 1.

The divergence is load-bearing:

• pstring is not bounded by max_string_length at all — it is bounded only by its length prefix and the buffer (~64 KB) 1.
• Nothing in src/output/ enforces a render-time truncation; concatenate_messages in src/lib.rs assembles raw rule-match strings without any output-length or newline guard 2.

The Gap: Escaping Is TTY-Gated to Symlink Targets Only#

The only terminal-control-byte escaping in the codebase is render_symlink_target() in src/cli/symlink.rs. It is:

• TTY-gated: escaping fires only when stdout is an interactive terminal (escape_control_bytes flag), preserving byte-for-byte parity with GNU file for captured/piped output 3.
• Symlink-target-only: the function is called from classify_symlink() for the "symbolic link to ..." and "broken symbolic link to ..." descriptions. Magic-rule descriptions never pass through it.

The escape set covers three families 4:

No equivalent function guards the format_magic_message() → concatenate_messages() pipeline that produces detection output from magic rules 2 5.

Related: /w Anchor Off-By-One (Same Issue)#

Issue #498 also documents that flagged_string_bytes_consumed advances by the walked text length (after greedy whitespace consumption by /w) rather than the pattern length 1. GNU file advances by pattern length. This is a separate bug filed under the same issue and affects relative-offset children of /w-flagged string rules.

Affected Output Path#

file content
  → read_string / read_pattern_match (scan-window bound: max_string_length)
  → format_magic_message() [src/output/format.rs] ← no escaping
  → concatenate_messages() [src/lib.rs] ← no escaping
  → stdout ← raw control bytes

The symlink path has escaping; the magic-rule path does not.

Current Status#

Issue #498 is open and unfixed as of 2026-09-21. The fix plan is documented in docs/plans/2026-09-21-1127-fix-bounded-inert-magic-descriptions-plan.md (not yet found in the main branch at time of writing). Three defects need to be addressed together 1:

1. Apply a 127-byte/first-newline render cap separate from the scan-window cap.
2. Fix flagged_string_bytes_consumed to advance by pattern.len(), not the walked length.
3. Apply TTY-gated control-byte escaping to magic-rule descriptions, mirroring render_symlink_target().

Related tracking: #382 (ascmagic text-class append), #379 (strength ordering), #304 (prior max_string_length enforcement gap).

Key Source References#