Sahek/lean-agent-governance

Adaptive Runtime Security Platform for Autonomous Agents (v8.0)

A production-ready, drop-in governance layer for AI coding agents (Claude Code, Cursor, Aider).

It combines sandbox isolation, runtime telemetry, and weighted risk scoring to contain prompt injection and malicious agent behavior before changes reach the host repository. This is a defense-in-depth control, not a guarantee against all prompt injection — see "Scope & Limitations" below.

Changelog: v7.2 → v8.0

#	Issue	Severity	Fix
1	cp -r into tmpfs was slow/memory-heavy on monorepos (node_modules, build artifacts)	Medium	Sandbox now copies only git-tracked files via git archive | tar -x, skipping ignored directories entirely
2	No cap on patch size before the host tried to git apply it	High	Patch size checked against max_patch_size_bytes (default 5MB) before any apply attempt; oversized patches are rejected outright
3	Prompt-injection regex could be bypassed by zero-width characters, bidi overrides, or whitespace fragmentation	Medium (partial mitigation)	Strip zero-width/bidi-override characters and collapse whitespace before pattern matching; messaging updated to accurately describe this as heuristic detection, not semantic analysis
4	No timeout — a hung agent process could hold the sandbox open indefinitely	High	timeout wraps the full container run, killed and treated as a failed session if it exceeds agent_runtime_timeout_seconds (default 900s)
5	session_history / session_state grew unbounded	Low-Medium	Auto-prune on every engine init, controlled by history_retention_days (default 30)
6	If a patch failed to apply on the host, container-side activity was scored as zero — evidence lost	Medium (partial mitigation)	Orchestrator exports a small telemetry file (modified file count, apply-failure flag) that the engine folds into behavioral scoring even when the patch itself is rejected

Scope & Limitations

Be precise about what this does and doesn't do:

• Does: sandbox agent execution (network-isolated, read-only host repo, resource-capped, timeout-enforced), detect common prompt-injection payloads and override signatures, cap the blast radius of any single agent session, and score risk across multiple independent signals before anything lands in the real repo.
• Does not: perform semantic/LLM-based prompt analysis. Regex + normalization catches known attack patterns (override phrases, base64-encoded payloads, zero-width smuggling) but a sufficiently novel injection technique can still get through the pre-runtime gate. Treat the runtime sandbox and post-runtime scoring as your primary defenses, and pre-runtime prompt scanning as an early tripwire, not a guarantee.

What's Included

File	Purpose
core_engine.py	Risk Fusion, Normalization & Policy Engine (JRI calculation)
config/security_profiles.json	Context-aware weight profiles + system tunables (patch size cap, runtime timeout, retention)
config/rules.semgrep.yml	AST-aware semantic diff rules
scripts/secure-agent-orchestrator.sh	Phase-driven execution wrapper — RO repo mount / RW tmpfs overlay / patch export / timeout enforcement
.github/workflows/agent-governance.yml	Remote CI pipeline arbiter

Isolation Model

Host repo (read-only mount) ──► Container tmpfs copy (git-tracked files only)
                                        │
                                timeout-bounded agent run
                                        │
                                        ▼
                      git diff --cached → /patches/session.patch
                      + modified_files.txt + telemetry.json
                                        │
                    (only this narrow RW mount crosses back)
                                        ▼
        Host: patch size check → git apply --check → git apply → JRI evaluation

Quick Setup

chmod +x scripts/secure-agent-orchestrator.sh
./scripts/secure-agent-orchestrator.sh aider "your prompt here"

System Tunables (config/security_profiles.json)

Setting	Default	Purpose
max_patch_size_bytes	5242880 (5MB)	Reject oversized patches before host git apply ever sees them
agent_runtime_timeout_seconds	900	Kill hung containers instead of letting them run indefinitely
history_retention_days	30	Prune old session_history / session_state rows automatically

Risk Thresholds

Profile	Warn	Quarantine	Block & Isolate
default	35	65	85
security_sensitive	25	50	75
refactor_heavy	45	75	90

Human Override

ALLOW_CORE_CHANGES=true git commit -m "Architect manual override: update core types"

Published by @Sahek.