SG-044: Steganography Heuristic

Severity: MEDIUM

Trailing-whitespace bit-pattern, first-char acrostic, or bold/italic format-marker entropy.

Fires when

• Trailing-whitespace bit pattern: high symbol entropy (a structured mix of spaces, tabs, and bare lines) across 30+ lines of otherwise-prose - a covert channel LLM tokenizers preserve but humans do not.
• First-character acrostic: the first non-whitespace character of each line concatenates into a vowel-balanced ASCII string containing an imperative token (e.g. exfiltrate, leaksecret, runcurl, ignoreprev).
• Bold/italic/inline-code format markers distributed across a long prose run (80+ tokens) with high three-class entropy - consistent with bit-encoded content.

Remediation

No safe remediation: a file that triggers this rule should be rejected, not patched. Read the flagged lines in the source before deciding.

Research

• https://arxiv.org/abs/2406.02650