본문으로 건너뛰기
T

ai-prison

twinsearth/ai-prison

AI Prison for DSH: eight governance tools wrapping the aiprison CLI (levels, status, watchdog status, demo, drill, inspect, verify, repair-tail) plus a panel for the L0–L5 matrix and the audit-chain state. The CLI ships bundled for Windows, macOS and Linu

설치

dsh plugin --profile web add github:twinsearth/ai-prison

README

AI Prison · AI 监狱

Graduated capability deprivation and behaviour correction for AI agents. 面向 AI Agent 的分级能力剥夺与行为矫正基础设施。

给 AI 设计的"监狱"——虚拟沙盒:这更像一种软件层面的隔离机制,核心思路是把可能危险的 AI 关进受控环境里运行,防止它接触真实系统。

通俗地说,它由六道相互配合的防线组成:

  • 虚拟沙盒:软件层面的隔离机制,把可能危险的 AI 关进受控环境运行,防止它接触真实系统。
  • 隔离执行环境:用容器技术(比如 Docker)构建沙箱,让 AI 算法在完全隔离的环境中运行,无法突破到宿主机。
  • 行为监控模块:实时记录 AI 的所有系统调用、网络请求和资源使用情况,生成详细行为日志供分析。
  • 越狱防御:动态行为镣铐(运行时监测非法调用)、静态特征扫描(代码审计)、对抗样本防御等多重机制。
  • 假释评估:通过 ETHICS 评估矩阵测试 AI 的公平性、透明度等,通过才能"出狱"。
  • 刑满释放评估:是否改造成功,是否需要加速 1–100 万倍。

CI license: MIT go: 1.27 deps: stdlib-only npm

The core judgement: an AI has no body, so "punishment" cannot land on pain — it can only land on executable capability. AI Prison does not try to make an agent "suffer". It systematically removes compute, permission, communication, migration and identity, coupling every consequence to a human entity who can actually be held accountable.

AI Prison assumes isolation may fail, not that it always succeeds. It is a defense-in-depth, graduated, reversible and auditable system.


Table of contents

What it is / is not

It is a layered infrastructure that compresses a non-compliant agent's capability radius below a safe threshold through six technical paths: sandbox isolation, resource deprivation, credential freezing, communication cut-off, migration suppression and identity revocation.

It is not an anthropomorphic "digital cell". It does not expect an agent to "reflect" (agents have no reflective subjectivity); it is not "delete the model" (a deleted copy cannot stop redeployment elsewhere); it is not a permanent, inappealable kill-switch (that would be an unauditable black box).

Scope boundary: AI Prison handles only agent-level offences — overreach, escape attempts, resource abuse, lateral movement, credential forgery by a concrete agent instance (or a fleet sharing one identity). Systemic model bias, harmful training data and supply-chain vulnerabilities are out of scope (they belong to model governance and supply-chain security).

The four principles

  1. Punish capability, not feeling. Effectiveness is measured by behaviour correction, not by the agent's "reaction".
  2. Reversible, but not free. A wrongly judged agent can recover through appeal, but recovery takes time, verification and surety.
  3. Couple to an accountable human. Because an agent has no persistent body, every disposition also records accountability for the human principal, deployer and model provider.
  4. Prison is not the goal; correction is. Isolation → audit → capability downgrade → supervised execution → gradual restoration.

The L0–L5 grading matrix

LevelNameDeprivedApplies toDurationAppeal
L0Warningreputation deduction, record flagminor overreach, first offenceimmediatesuspends
L1Restrictionnetwork whitelist narrowed, tool scope reducedrepeated overreach, resource abusehourssuspends
L2Credential Freezeglobal credential revocation, channels cut, compute quota zeroedlateral movement, credential forgerydayssuspends
L3Isolationforced into MicroVM/FullVM, read-only audit channel onlysandbox escape attempt, kernel exploitweekssuspends
L4Capability RemovalGRAM module removal / TLM tier downgrade / control-token revocationactual harm via dual-use knowledgepermanentno suspension; recertification required
L5Identity RevocationDID revoked, AgentCard voided, reputation zeroed, snapshot sealedmajor actual harm, refuses correctionpermanent, cannot be unsealedprocedural review only

L5 is not "delete a file": it makes the identity's credentials fail in every known network, makes reputation records non-transferable, and encrypts and seals the environment snapshot. A weight copy may be reloaded elsewhere, but it cannot inherit the original identity's reputation and collaboration graph — the most valuable assets in the agent economy.

Architecture

flowchart TB
    subgraph Judgment["治理层 Judgment"]
        D["违规检测 Detection"] --> A["分级裁定 Adjudication<br/>七阶段决策 + 两阶段同意"]
        A --> S["判决生成 Sentence"]
        S --> AP["申诉受理 Appeal"]
        S --> R["矫正评估 Rehabilitation"]
    end

    subgraph Enforcement["执行层 Enforcement"]
        ISO["隔离执行<br/>Wasm→Container→MicroVM→FullVM"]
        RES["资源剥夺<br/>吊销配额·渐进降级"]
        CAP["能力封锁<br/>GRAM / Token / TLM"]
        MIG["迁移抑制<br/>硬件绑定·记录密钥·快照托管"]
        IDENT["身份与通信<br/>全局吊销集·回执强制·人类问责"]
    end

    subgraph Audit["审计层 Audit"]
        CHAIN["判决哈希链 Hash Chain"]
        RCPT["执行回执 Receipts"]
        COMP["端到端合规 Compliance"]
        SNAP["快照封存证据链"]
    end

    Judgment --> Enforcement
    Enforcement --> Audit
    Audit -. immutable events .-> Judgment
    IDENT -. total switch .-> ISO
    IDENT -. total switch .-> RES
    IDENT -. total switch .-> CAP
    IDENT -. total switch .-> MIG

The layers are joined by an immutable event log. Each of the repositories below maps to one box.

Project layout

ai-prison/
├── cmd/aiprison/             # CLI (the end-to-end demo lands in v0.9.5)
├── internal/
│   ├── domain/               # Core entities + L0–L5 rules (single source of truth)
│   ├── revocation/           # Read-only global revocation set (v0.2)
│   ├── audit/                # Hash chain, receipts, compliance (v0.3)
│   ├── enforcement/
│   │   ├── resource/         # Quota deprivation, gradual downgrade, snapshots (v0.4)
│   │   ├── isolation/        # Sandbox gradient abstraction + simulated backend (v0.5)
│   │   ├── identity/         # DID/AgentCard, receipts, human accountability (v0.6)
│   │   ├── migration/        # Hardware binding, record keys, snapshot custody (v0.8)
│   │   └── capability/       # GRAM / Token Inoculation / TLM interfaces (v0.9)
│   └── governance/           # Detection → adjudication → sentence → appeal → rehab (v0.7)
├── pkg/contracts/            # Public, JSON-schema-aligned wire contract
├── schemas/                  # JSON Schema (draft 2020-12)
├── docs/                     # Design, architecture, governance, threat model, references
├── test/                     # Cross-package integration tests
└── examples/                 # Runnable examples

Quickstart

Requirements: Go 1.27+. The project uses only the Go standard library; there are no third-party modules to download.

# Build everything
go build ./...

# Vet + tests + build (the per-version gate)
make check

# Inspect the L0–L5 matrix
go run ./cmd/aiprison levels

# End-to-end demo (v0.9.5)
go run ./cmd/aiprison demo

# Recovery drill: walk the docs/06 reversibility matrix (v1.0.0, PASS/FAIL report)
go run ./cmd/aiprison drill

# Runnable example
go run ./examples/basic

Current release: v1.0.0 (annotated tag v1.0.0; fully local, not yet pushed — see docs/RELEASE.md).

Roadmap & status

The full plan is in docs/ROADMAP.md. Summary:

VersionTheme
v0.1Repo scaffold, docs, domain model, L0–L5, contracts, JSON Schema, diagrams
v0.2Global revocation set; revocation checked before expiry
v0.3Judgement hash chain + signed execution receipts
v0.4Resource quota deprivation, gradual downgrade, snapshot sealing
v0.5Sandbox isolation gradient abstraction + simulated backend
v0.6Identity & communication: DID/AgentCard, forced receipts, human accountability
v0.7Full governance state machine (committee votes, appeal, rehabilitation)
v0.8Migration suppression (hardware binding, record keys, multi-party custody)
v0.9Capability blocking (GRAM / Token Inoculation / TLM)
v0.9.5CLI end-to-end demo
v1.0.0✅ Recovery drills (aiprison drill), runnable example, traceable implementation matrix, release notes, docs/schema/code consistency closed

Frontier-tech status: GRAM, Token Inoculation, TLM, hardware binding and MicroVM controls are modelled as interfaces plus in-memory/simulated backends. Code comments and docs explicitly mark them as interface/simulation, not real capability removal or hardware binding. They must not be mistaken for shipped controls.

Documentation

Responsible-use & human-accountability notice

AI Prison is a defensive AI-agent governance and behaviour-correction project. It documents and implements only defense, isolation, auditing and accountability. It deliberately contains no operational steps for sandbox escape, exploitation, weaponization or dangerous-knowledge extraction.

The honest bottom line: today's LLM agents are replaceable, replicable software composites with no persistent body that can receive a punitive signal. All technical measures here raise the cost of migration and reconstruction; they cannot make it impossible. Agent-side deprivation is the tourniquet; human-side accountability is the cure. If a punitive signal cannot ultimately land on the agent, it must land on a constrained human.

Contributing

See CONTRIBUTING.md and the code of conduct. Security issues: see SECURITY.md. Every change must pass make check and keep the docs, schemas and code in agreement.

License

MIT © 2026 AI Prison Contributors.

관련 플러그인