dsh-defend
Security plugin for DeepSeek Harness that detects prompt injection, jailbreaks, secret leaks, and risky recursive-delete commands.
Install
$ dsh plugin --profile web add "github:PerryLink/dsh-defend#main"Plugin Overview & Capabilities
AI-assisted organization based on the public repository snapshot. The content must be grounded in source evidence and does not replace compatibility or security verification.
Key Capabilities
- Scans inbound user messages for prompt-injection, jailbreak, and secret-leak patterns at agent/pre-step.
- Scans tool arguments before execution and tool results after execution.
- Applies configurable allow, ask, or block actions by detection family; critical secrets block by default.
- Guards reviewed shell tools against unsafe recursive-delete commands, while allowing documented dry-run markers.
- Provides a defend_report tool and /defend command with aggregate and recent-match summaries without matched text.
- Writes sanitized defend/detection session audit events and keeps a bounded in-memory report buffer.
- Runs locally with declared network:none permission and no subprocesses.
Useful For
- Add deterministic prompt-injection and jailbreak screening to a DeepSeek Harness profile.
- Require approval or deny suspicious tool arguments before execution.
- Prevent recursive deletion outside explicit session-workspace paths.
- Detect secret-like content in messages or tool output without recording matched values.
- Review recent detection totals through the defend_report tool or /defend command.
Who It Fits
- DeepSeek Harness users seeking local security-policy interception.
- Teams operating agents with shell or terminal tools.
- Plugin administrators who want configurable detection actions and session audit records.
Documented Limitations
- Detection is rule- and heuristic-based; novel phrasing, lookalike-Unicode encodings, and multi-step attacks can evade it.
- The plugin does not make model-level intent judgments because it does not call a model.
- agent/pre-step rejection cannot provide a reason to the model because the referenced seam has no reason field.
- Scanning is capped at the configured maxScanChars value per interception, scanning the head of content.
- On hosts whose Session.append predates the ignorable marker, audit logging is disabled by default after capability detection to avoid unresumable-session risk; users can opt into unmarked audit logging.
DSH Compatibility
Version-specific runtime evidence collected by DSH Plugin. A missing result means we have not tested that combination yet.
Security Signals
Objective signals discovered from package metadata and source inspection. These are not a guarantee that a plugin is safe.
Source & Registry Notes
Public provenance, Registry classification, and the latest source check for this entry, kept separate from runtime verification.
- Source repository
- PerryLink/dsh-defend
- Registry source
- GitHub · dsh-plugin topic
- Registry classification
- Plugin
- Source checked
- 087ce8d · 2026-09-19
This project is independently indexed from public source information. DSH Plugin is not affiliated with DeepSeek or the plugin author. Always check the author repository before installation.
Repository Activity
- GitHub stars
- 16GitHub stars
- Forks
- 1
- Open issues
- 1
- Last commit
- 2026-09-19
- Last release
- 2026-09-18
Related DSH Plugins
Ranked by overlapping capabilities, use cases, plugin type, categories, and DSH profile.
Second-model AI auto-review for DeepSeek Harness approval requests with fail-closed safety and session audit.
View plugin →Composable three-tier memory control plane for DeepSeek Harness with runtime context, searchable documents, and pluggable long-term memory.
View plugin →Sandbox-first automatic permission policy and intelligent risk classification for DeepSeek Harness.
View plugin →Persistent multi-model Agent teams and bounded DAG workflows with GUI controls for DeepSeek Harness.
View plugin →