deglyph parses untrusted input by design: a binary under analysis may be malformed or hostile. The container parser (LIEF) and disassembler (Capstone) run over attacker-controlled bytes. The loader is written to fail with a clean error rather than crash, and the per-instruction and per-section scanners catch and continue so one bad region cannot abort a whole-image pass.

deglyph does not execute the binary it analyzes. It reads the file, parses its structure, and disassembles its code; it never runs the target.

The optional AI assistant sends data off the machine. When a question is asked in the Assistant tab, the selected function's disassembly, the conversation, and whatever the assistant's read-only tools read from the binary are sent to the configured model provider: the Anthropic API by default, any OpenAI-compatible endpoint chosen in the provider settings, or the hosted deglyph service when a deglyph login token is stored. A local provider (Ollama, LM Studio) keeps that traffic on the machine. Nothing is sent until a question is asked and a provider is configured. Do not use the assistant on binaries whose contents may not leave your environment.

CVE lookups (deglyph scan --cve) send the detected package URLs to osv.dev. They are opt-in and off by default; --offline forces the no-network path. No other feature contacts the network.

Report suspected vulnerabilities privately through GitHub: open the repository's Security tab and choose Report a vulnerability (Private Vulnerability Reporting). Please include:

• the deglyph version (deglyph --version) and platform,
• a description of the issue and its impact,
• a minimal input file or steps that reproduce it.

Please do not open a public issue for a security report. A maintainer will acknowledge the report and coordinate a fix and disclosure timeline.

deglyph is pre-1.0. Security fixes target the latest release and the main branch.