dsh-safeguard

In DeepSeek Harness, the model invokes real tools like bash, curl, and SQL. A single misfired rm -rf /, git push --force, or an AKIA key leaking into tool arguments can cause irreversible damage. dsh-safeguard hooks into the tools/pre-execute waterfall and short-circuits the call before execution, so the intercepted side-effect never happens.

It covers two risk categories out of the box: fifteen dangerous-command rules (rm -rf, force push, mkfs, dd, curl|sh pipes, SQL DROP/DELETE/TRUNCATE, and more) and six secret-leakage patterns (AWS AKIA, sk-, GitHub PAT, PEM private keys, Slack tokens). Rules are fully configurable via extraPatterns, allowList exemptions, and enableDanger / enableSecrets toggles, and two helper tools (guard_list, guard_check) let the model perform programmatic self-checks at runtime.

To be clear, this is a regex-based heuristic guard: it does not verify signatures, validate key authenticity, or provide process or filesystem sandboxing. It is best suited for admins who run DSH in production pipelines and want a low-cost interception layer for model tool calls, to be used alongside dedicated secret scanners and OS-level sandboxes rather than as a replacement for them.