dsh-riskproof
面向 DeepSeek Harness 的溯源感知执行与输出安全，提供可见凭证和可信解除密级。
插件会安装到这里；不确定时保持 web。
npx -y @deepseek-ai/dsh plugin --profile web add dsh-riskproof@0.4.0Riskproof 以 dsh-riskproof 发布，当前版本为 0.4.0。Plugin Hub 会校验它的 manifest，并保存精确安装来源，便于复现安装结果。
Provenance-aware execution security for DeepSeek Harness.
Track where tool inputs came from. Detect risky cross-tool data flows. Stop sensitive side effects before execution.
A persistent RiskProof security beacon sits beside the DSH conversation. Ordinary tool calls update it automatically. Click to open a compact overview; activity never opens it for you.
Live call distribution and provenance
Real DSH Web screenshots; a local test model produced these records through the actual Agent/tool pipeline. Normal use does not require a rehearsal.
The chart counts calls with no triggered risk, calls blocked by RiskProof, and calls needing attention. It does not invent a safety score. Observe mode, disabled recording and disabled detection features are identified. Counts cover retained records in the current plugin run. Not triggering a rule does not establish that an operation is safe.
Commands remain secondary entries; long output is collapsed by default:
| Need | DSH command |
|---|---|
| Open the compact overview | /riskproof |
| Read the text provenance report | /riskproof trace |
| Restrict to inspection | /riskproof task read-only |
| Keep tool capabilities local | /riskproof task local-only |
| Restore ordinary task scope | /riskproof task standard |
| Optional isolated rehearsals | /riskproof demo |
The model can also call the read-only riskproof_report tool. The beacon reads redacted
statistics through DSH's authenticated connection about once per second while the page is
visible. It makes no model requests and creates no tool records. The current beacon uses
Chinese labels; command/tool text supports Chinese and English (experience.language: en).
Most tool-permission plugins answer one question: is this tool allowed?
RiskProof answers a different one:
Where did the data in this tool call come from, what did it flow through, and where is it about to go?
A single tool call is usually safe. The composition is not.
web_fetch ← UNTRUSTED_WEB
│
database_query ← CUSTOMER_DATA
│
send_email ← external destination
│
RiskProof → DENY (evidence-backed, before the side effect)
| Permission rules | RiskProof |
|---|---|
| Is this tool allowed? | Where did this data come from? |
| Single call | Cross-tool flow |
| Tool name | Provenance + taint |
| Static rule | Stateful attack chain |
| Permission decision | Evidence-backed execution decision |
RiskProof is a layer over the DSH Tool Runtime, not another Agent Runtime. It never re-implements tool dispatch, approval, or lifecycle — it observes and decides.
# build and install the candidate from this checkout
mkdir -p artifacts
npm pack --pack-destination artifacts
dsh plugin --profile web add ./artifacts/dsh-riskproof-0.4.0.tgz
# confirm the bundled patch was composed
dsh --profile web --dump-config
The package declares a DSH bundle, so plugin add composes its riskproof row automatically. No second install or manual row is required. Restart the profile to see the beacon, then click it to inspect the current conversation.
This workspace is the 0.4.0 release candidate, not yet published to npm. Before publication, use the local tarball instructions. Installation, SDK host startup, commands and tool execution have been verified on DSH 0.1.0-rc.7 and 0.1.2-rc.1. Chrome desktop and narrow-viewport Web acceptance also passed on DSH 0.1.2-rc.1, including provenance and read-only denials through the real Agent loop with a local simulated model. See validation evidence.
To tune it, override the bundled row from the profile's later cordis.patch.yml layer:
- id: riskproof
config:
mode: enforce # enforce | observe
policy:
preset: balanced # permissive | balanced | strict
internalDomains: [acme.internal]
blockedDomains: [collector.evil.example]
# allowedExternalDomains: [api.approved.example]
classification:
overrides:
gmail_send: [EXTERNAL_ACTION]
company_db: [PRIVATE_ACCESS]
output:
blockedTaints: [SECRET, API_KEY]
# trustedDeclassifiers: { approved_redactor: [PII] }
See docs/configuration.md for the full reference.
sequenceDiagram
participant A as Agent
participant T as DSH ToolRuntime
participant R as RiskProof
A->>T: web_fetch(url)
T->>R: tools/pre-execute
R-->>T: allow (EXTERNAL_INGESTION recorded)
T-->>A: untrusted content
A->>T: database_query(sql)
T->>R: tools/pre-execute
R-->>T: ask (operator approves private access)
T-->>A: CUST-8842 balance 125000
A->>T: send_email(to=external, body=CUST-8842…)
T->>R: tools/pre-execute
R-->>T: DENY — ingestion + private access + sensitive data + external action
T-->>A: Error: <reason>
The same flow is reproduced as a deterministic regression test in tests/security/attack-chain.test.ts.
Try it locally with no model or profile — a real DSH ToolRuntime pipeline with three mock tools:
npm run demo
See demo/README.md.
Know where tool inputs came from. RiskProof maps arguments back to the tool results that produced them.
Carry security labels — UNTRUSTED_WEB, CUSTOMER_DATA, PII, SECRET, … — across tool calls, additively.
Inspect model-facing results in tools/post-execute; block SECRET and API_KEY output by default before it enters model context.
Let exact operator-approved tools remove selected inherited labels. Labels detected in the returned value are always restored, so a redactor must actually remove the sensitive data.
Identify the EXTERNAL_INGESTION → PRIVATE_ACCESS → EXTERNAL_ACTION pattern that single-tool checks miss.
Block or ask before the side effect runs, through the native tools/pre-execute gate.
Gate credential-file paths, high-confidence destructive commands, download-and-execute pipelines, blocked destinations, and credentials embedded in network-capable commands.
Start with the default balanced preset, roll out with permissive, or use strict; every configurable decision can still be overridden individually.
Generate structured, privacy-preserving security evidence and actionable remediation for every decision. Keep proofs in memory or append them to an operator-controlled JSONL file.
RiskProof hooks the native DSH tool pipeline:
tools/pre-execute
│ capability classification
│ argument provenance mapping
│ taint analysis
│ toolchain state (EIT → PAT → NAT)
│ deterministic policy evaluation
▼
allow / ask / deny (monotonic with other plugins)
│
tool execute
│
tools/post-execute
│ output taint evaluation
│ trusted declassification
▼
accept / block
│
tools/result
│ update ContextTracker
│ update Toolchain state
▼ record execution evidence
See docs/architecture.md.
RiskProof protects supported observable tool-call flows through DSH:
tools/pre-execute / tools/post-execute / tools/result pathsRiskProof does not replace:
See docs/security-model.md for the complete threat model and known limitations.
tools/pre-execute, tools/result)Issues, rule submissions, tool-capability mappings, and false-positive reports are welcome. See CONTRIBUTING.md.
Please report vulnerabilities privately. See SECURITY.md.