dsh-promptwall
Local prompt-injection and secret-exfiltration firewall for DeepSeek Harness 适合需要在本地对输入输出进行安全扫描、拦截可疑注入和泄露的用户。
インストール
$
npx -p @deepseek-ai/dsh dsh plugin --profile web add github:Chhlafiu4312/promptwallドキュメント
README 全文を読む ↗Configuration
| Field | Default | Purpose |
|---|---|---|
enabled | true | Register the tool and policy hooks. |
injectionAction | sanitize | monitor, sanitize, or block suspicious output. Dangerous and truncated output still fails closed. |
suspiciousThreshold | 30 | Score that produces a suspicious verdict. |
dangerousThreshold | 70 | Score that produces a dangerous verdict. |
maxScanChars | 250000 | Maximum UTF-16 code units inspected per prompt or credential-bearing string; incomplete inspection fails closed. |
maxJsonDepth | 256 | Maximum canonical tool-result nesting depth; exceeding it fails closed. |
maxJsonNodes | 100000 | Maximum canonical JSON values inspected per tool result; exceeding it fails closed. |
inspectToolOutputs | true | Inspect post-execution output automatically. |
trustedTools | promptwall_scan | Exact tool names exempt from automatic reinspection. |
egressAction | ask | off, ask, or deny for secret-like egress arguments. |
egressToolPatterns | common network names | Case-insensitive patterns identifying egress-capable tools. |
rules | [] | Additional deterministic injection rules. |
secretPatterns | [] | Additional credential patterns. |
The complete default composition is in cordis.patch.yml. Custom rules are JavaScript regular-expression sources and should be reviewed like code.