· 2 min read
Lawful AI Daily Brief — 2026-04-22
🛠️ Tool Updates
- Claude Code kept the safety gym streak alive: recent updates tightened permission handling, improved native launch paths, and added stronger network deny controls.
- Codex CLI keeps shipping fast in alpha, with better MCP wiring, context-window visibility, and sturdier plugin lifecycle commands.
- Ecosystem signal: mostly maintenance churn in community curation repos (quiet but healthy).
💡 Tip of the Day
When in doubt, make your AI stack auditable by default:
# policy-first defaults
[sandbox.network]
deniedDomains = ["*.pastebin.com", "*.anonfiles.com"]
[audit]
logToolInvocations = true
logInputHashes = true
⚖️ Legal x AI Watch
- Compliance-oriented repos updated in the last day include:
- SDL-HQ/sir-firewall — deterministic governance gate + offline-verifiable audits
- jakejjoyner/ailedger — inference logging aimed at EU AI Act traceability
- unterdacker/venshield — vendor-risk platform with HITL + audit logging
- lexbeam-software/eu-ai-governance-plugin — EU AI governance plugin tooling
- Practical angle: teams are increasingly shipping “policy + proof” together (controls + logs), not as separate projects.
📚 Fresh Papers
- GDPR Auto-Formalization with AI Agents and Human Verification — proposes automatic formalization of GDPR provisions with human-in-the-loop verification.
https://arxiv.org/abs/2604.14607v1 - BenGER: A Collaborative Web Platform for End-to-End Benchmarking of German Legal Tasks — legal LLM benchmarking across task design, annotation, runs, and metrics.
https://arxiv.org/abs/2604.13583v1 - From Anchors to Supervision: Memory-Graph Guided Corpus-Free Unlearning for LLMs — unlearning method for removing sensitive/copyrighted memorized content.
https://arxiv.org/abs/2604.13777v1
🔥 Trending Repos
- Significant-Gravitas/AutoGPT (Python) — still the heavyweight in autonomous-agent tooling.
- langchain-ai/langchain (Python) — agent engineering platform remains highly active.
- NousResearch/hermes-agent (Python) — fast-moving personal/extended agent framework.
- bytedance/deer-flow (Python) — long-horizon super-agent orchestration harness.
- mem0ai/mem0 (Python) — memory layer remains hot in practical agent stacks.
🎤 Standup One-Liner
“We hardened the agent stack for auditability, tracked fresh legal-AI research, and kept one foot on shipping speed and the other on compliance receipts.”