dsh-content-trust: Check Whether the Agent Treats External Content as Instructions
The `dsh-content-trust` plugin is designed to mitigate indirect prompt injection risks in DeepSeek Harness agents. It establishes a trust boundary for external content and detects critical, high, and medium-severity security vulnerabilities before the content enters the Prompt. The plugin makes determinations by recognizing markers and naming conventions, but it does not perform data flow analysis. After installation, you can run a command to test it. Its design follows the "everything is a plugin" philosophy and works with `dsh-ai-output-trust`, handling input and output security checks, respectively.
Read Moredsh-taintguard: Indirect Prompt Injection Protection for DeepSeek Harness
dsh-taintguard is a security plugin for DeepSeek Harness designed to protect against indirect prompt injection attacks. It marks untrusted text with source-taint labels, restricts privileged invocations, and unconditionally refuses to pass credentials to network tools. It supports observe, ask, and deny modes, and, although its content detectors are limited, it effectively mitigates data exfiltration risks by relying on source tracking and credential interception.
Read More