deepsec is a security tool. We take vulnerabilities in deepsec itself seriously and ask that you give us a chance to fix them before they become public.

For vulnerabilities in deepsec — the scanner engine, the AI investigation pipeline, the bundled matcher pack, the AST bridge, the compliance reporter, or any first-party CLI — please do not open a public GitHub issue. Instead, use one of:

• GitHub Security Advisories (preferred): https://github.com/noeljackson/deepsec/security/advisories/new
• Email: security@noeljackson.com

PGP available on request.

Include:

• A minimal reproduction (a single command, a fixture file, or a patch is ideal).
• Affected versions / commits — deepsec --version plus the branch/tag.
• Your assessment of severity and blast radius.
• Whether you've already disclosed publicly anywhere (please don't, but we want to know if you have).

Vulnerabilities in the deepsec project itself:

• CLI / engine: command injection, path traversal, or unsafe deserialization in deepsec's own code paths.
• Matcher engine: a malicious matcher TOML that escapes into arbitrary code execution, sandbox escape, or denial-of-service against the scanner.
• AI backends: prompt injection that exfiltrates project source beyond the configured backend, or that causes deepsec to take destructive actions outside the user's scope.
• AST bridge (wazero / tree-sitter WASM loading): a crafted grammar that escapes the wazero sandbox.
• Compliance reports: signed-manifest forgery, signature downgrade, or HMAC key extraction.
• CI integrations: pre-commit hook or GitHub Action that leaks the scanned repo's secrets via deepsec's own action.

• Vulnerabilities deepsec finds in user code. That's the product working — file a GitHub issue or follow the user's own security channel.
• Vulnerabilities in upstream Go modules, tree-sitter-* npm packages, wazero, or other dependencies. Please file them upstream; we'll pick up the fix on the next bump.
• Misconfiguration of deepsec by the user (API keys leaked into logs, public CI output, etc.). The product can help mitigate via --max-cost-usd, --record, sensitive-log matchers — but the responsibility is the user's.
• DoS via expensive AI calls. Already addressed by --max-cost-usd and deepsec spend — file a feature issue if you want stronger controls.

• Acknowledgement: within 3 business days.
• Initial triage: within 7 business days.
• Fix or workaround: depends on severity. Critical vulns get an out-of-band patch release; high/medium go into the next regular release; low get a normal-priority PR.
• Disclosure: coordinated. Once a fix is shipped and most users have had time to update (typically 30 days), the GitHub Security Advisory becomes public. Reporters are credited unless they prefer otherwise.

The short version: deepsec is run by a developer or CI bot, against source code that may be untrusted, calling out to an LLM provider the user configures. The trust boundaries:

A fuller threat-model document lives at docs/threat-model.md (TBD, issue #94 follow-up).

(Will list reporters of accepted vulns once we have any. None yet.)