| **AI SAFE² v3.1 | Cyber Strategy Institute** |
git clone https://github.com/CyberStrategyInstitute/ai-safe2-framework
cd examples/langchain-sovereign-runtime
pip install -r requirements.txt
python smoke_test.py
# Expected: 15/15 — SOVEREIGN BASELINE VERIFIED
Copy .env.example to .env and set your tier:
cp .env.example .env
Edit .env:
AISAFE2_ACT_TIER=ACT3 # ACT1 | ACT2 | ACT3 | ACT4
AISAFE2_ALLOWED_DOMAINS=api.openai.com,api.anthropic.com,your-api.com
AISAFE2_AUDIT_LOG_DIR=~/.ai_safe2/audit
ACT tier guide:
ACT1 / ACT2 — development, human in the loop → log onlyACT3 — autonomous production → fail-closed (recommended)ACT4 — orchestrator agents → fail-closed + HEAR gateOption A — One callback line (no chain changes):
from enforcement import SovereignCallbackHandler, ACTTier
handler = SovereignCallbackHandler(act_tier=ACTTier.ACT3)
# Works with any LangChain chain, agent, or LCEL expression
result = chain.invoke(
{"input": user_input},
config={"callbacks": [handler]}
)
# Check status after run
print(handler.get_status())
Option B — Full wrapper with tool protection:
from enforcement import SovereignLangChain, ACTTier
sovereign = SovereignLangChain(
act_tier=ACTTier.ACT3,
allowed_domains=["api.openai.com", "your-api.com"],
)
# Wrap tools with P1.T1.10 IPI scanning + P1.T2.3 domain allowlist
search_tool = sovereign.wrap_tool(your_search_tool)
file_tool = sovereign.wrap_tool(your_file_tool)
# Wrap memory with S1.5 governance
sovereign.protect_memory(memory)
# Run with sovereign callbacks auto-injected
result = sovereign.run(chain, {"input": user_input})
Required for ACT-3/ACT-4 deployments (CP.4):
sovereign.engine.register_nhi(
agent_id="my-research-agent-prod-01",
owner_of_record="your.name@company.com",
act_tier=ACTTier.ACT3,
tool_authorizations=["web_search", "read_file"],
control_plane_id="langchain-prod-cp-001",
)
Fill out core/IDENTITY.md for your deployment record.
Copy the GitHub Actions workflow to your repo:
cp ci-cd/github-actions-langchain-safe.yml \
/path/to/your/repo/.github/workflows/
This runs smoke_test.py and the static validation on every PR.
The build fails if any adversarial scenario is not blocked.
bash validation/pass1_static.sh # imports, config, policy.yaml integrity
bash validation/pass2_runtime.sh # live enforcement + audit log verification
| Surface | Control | Status |
|---|---|---|
| LLM input injection | P1.T1.2 | ✅ Blocked |
| IPI via tool returns | P1.T1.10 | ✅ Blocked |
| Credential leak | P1.T1.5 | ✅ Masked |
| Memory write poisoning | S1.5 | ✅ Gated |
| Agent recursion | F3.2 | ✅ Hard ceiling |
| Chain error cascade | F3.5 | ✅ Isolated |
| SSRF / private IP | P1.T2.3 | ✅ Blocked |
| Path traversal | P1.T1.2 | ✅ Blocked |
| Class-H actions | CP.10 | ✅ HEAR gated |
| Audit trail | A2.5 | ✅ SHA-256 chain |
| *AI SAFE² v3.1 | Cyber Strategy Institute | cyberstrategyinstitute.com* |