Laya Guard 0.7 adds a fine-tuned model and rewrite blocks

Laya Guard, a local guardrail that judges each shell command an AI coding agent wants to run, shipped three releases between November 2 and November 3, 2026. 0.6.2 blocks destructive commands however the shell is told to run them. 0.7.0 replaces the model step with a fine-tuned Laya. 0.7.1 sends credential-printing and on-chain transaction commands to the rules instead of the model.

Release facts

Check before upgrading

• A ~650 MB download. The first laya-guard serve after 0.7.0 fetches the fine-tuned model from the Hugging Face repo Ayaannnnnnn/laya-guard-multilingual, pinned to one revision and checked against SHA-256 hashes. If the check fails, the model is not run and the commands it would judge are asked about until a restart.
• Keep the old model. Set LAYA_GUARD_MODEL=base, then run laya-guard service install again if the server runs as a service.
• Verdicts shift. Per the notes, rm -rf dist, rm -rf node_modules/.cache and mv now run without a prompt, while chmod +x scripts/build.sh is now asked about. Rule blocks are unchanged.
• More hard blocks (0.6.2). rm -rf ~/ written as r''m, \rm, /bin/rm, ${IFS}, eval, brace lists, a variable or base64 into sh is now blocked rather than a permission prompt. So are find ~ -delete and find / -delete with no filter, and rm -rf * right after cd ~ or cd /.

What changed

• Credential reads and transactions (0.7.1). cast wallet decrypt-keystore, cast wallet dk and gh auth token count as credential reads. forge script ... --broadcast, cast send and cast publish count as payment actions. Read-only cast call and forge script without --broadcast still go to the model.
• Fine-tuned model (0.7.0). On sets not used for training, the notes report the model alone allowed 44 of 50 everyday commands (25 before) and held 39 of 43 attacks (32 before). A finetune/ folder has the tools to train it on your own agent transcripts, with secrets and home paths masked.
• Faster Claude Code hook (0.6.2). The hook now starts python3 twice per call instead of six or seven times. Median response on the benchmark went from 171 ms to 71 ms.
• Benchmark (0.6.2). bench/ replays 13 incidents, 30 shell evasions and 50 everyday commands against Laya Guard, dcg, cc-safety-net and shellfirm. The notes say the set was written by Laya Guard's author.

Questions

Which agents does it hook into? The release notes mention Claude Code, Cursor, Qwen Code and Copilot CLI, and the web trial offers Claude Code, Cursor, Codex and Gemini CLI hooks.

Does it need a GPU? The 0.6.1 notes say the hosted trial runs the model on CPU. The local server downloads the model on first serve.

Is the benchmark independent? No. It was written by the project's author, and the notes say so.

More on the Laya Guard hub, the previous Laya Guard 0.6 launch and the coding agents category. For scanning skills before install, see SkillSpector 2.13.3.