返回资讯中心
外部精选
软件工程
#开发者实时流

Nono: Agent Runtime Security

Agent runtime security. Kernel-enforced boundaries for AI agents, with zero setup and zero latency.

Hacker News 最新pessimizer2026/10/17 · 2026/10/1711 分钟阅读

以下正文同步自 Hacker News 最新,版权归原站所有,已转换为易读排版。

Agent runtime security. Kernel-enforced boundaries for AI agents, with zero setup and zero latency.

Install

curl

$curl -fsSL https://nono.sh/install.sh | sh

Homebrew

$brew install nono

Debian · Fedora · Arch · Nix →

Quickstart

Find a package

$nono search pi↳nolabs-ai/pi

Run it, sandboxed

$nono run --profile nolabs-ai/pi -- pi

Created by the team behind Sigstore

The industry standard for software signing, used by PyPI, Homebrew, Maven, Google, GitHub, and NVIDIA.

Every tool call gets an ephemeral micro toolbox, scoped to only the files, network routes, arguments, and credentials it needs. Every decision is auditable.

How nono's ephemeral micro toolboxes work: a coding agent runs a tool invocation, for example gh issue view. The nono supervisor resolves and verifies the executable, evaluates argv policy for the caller and arguments, and on approval creates a fresh, invocation-scoped micro sandbox around the tool with only its selected capabilities: read-only workspace, no filesystem writes, no host secrets, bounded stdout and stderr, and network access only through the nono proxy. The sandbox receives a phantom GH_TOKEN; the real credential stays with the supervisor and never enters the child. The proxy validates the phantom token, evaluates HTTP method and path policy, injects the real credential at the boundary, and forwards over TLS to api.github.com. If an action falls outside the provided policy, the policy engine can route it to a human approval decision to approve, deny, or time out. Security-relevant events are hash-chained into an audit record sealed with a SHA-256 Merkle root, and the sandbox is destroyed when the invocation exits. Four scenarios: Allowed read — gh issue view runs, POST /graphql is allowed by the proxy, and output returns to the coding agent. Human approval — gh pr merge matches no argv rule, so the exact invocation pauses for a human decision and, once approved, executes inside a scoped sandbox. Argv denied — gh issue comment is refused at argv authorization; no sandbox or outbound request is ever created. L7 denied — gh api passes the broad argv rule, but the POST to the repository comments endpoint is denied at the proxy; nothing reaches GitHub.

AGENT SESSIONLIVEUUSERWhat changed in issue #1052?AGENTLet me check the issue.ghTOOL CALLGitHub · View issuenolabs-ai/nono · #1052TOOL RESULTIssue #1052 · Brokered tool isolationOpen · 14 commentsRESOLVED INVOCATIONgh issue view 1052 --repo nolabs-ai/nonoNONO · POLICY + CAPABILITY BROKER01EXECUTABLEgh · digest verified02ARGV POLICYissue viewALLOW03CAPABILITIESread-only · proxy-only04CREDENTIALreal token remains hereOUTSIDE PROVIDED POLICYHUMAN APPROVALapprove · denyHUMAN-IN-THE-LOOPNONO MICRO SANDBOXINVOCATION-SCOPEDPROCESS MODULE>_ ghPID 48291 · issue view 1052FILESYSTEMWORKSPACE · READWRITESDENIEDSTDIO1 MB · BOUNDEDNETWORKPROXY ONLYGH_TOKEN=nono_cd07…L7 PROXYINSPECTPOST /graphqlALLOWINTERNET · GITHUB APIapi.github.comTAMPER-EVIDENT AUDITresolveargv.allowspawncredentiall7.allowseal9a3b7c1d…6082

AGENT SESSION

LIVE

USER

What changed in issue #1052?

INTERNET · API.GITHUB.COM

TAMPER-EVIDENT AUDIT

resolve → argv → spawn → l7 → seal

MERKLE ROOT

9a3b7c1d…6082

Allowed: gh issue view ran inside a fresh invocation-scoped micro sandbox with a phantom GH_TOKEN. The proxy allowed POST /graphql and injected the real credential at the boundary; the audit record was sealed with a SHA-256 Merkle root and the sandbox was destroyed.

The agent never talks directly to the isolated tool. Requests, credentials, network traffic, stdio, and audit events cross the supervisor boundary.

nono sandboxes any terminal agent. Pull a signed profile from the registry and run — no wrappers, no rewrites.

At scaleEngineers at some of the world’s largest tech companies run nono to protect their agents.

It scales from a single laptop to fleets of agents without changing the model — the same composable policy, per-tool sandboxing, credential protection, and fully auditable execution.

Composable policy

Scalable JSON profiles that compose across teams and version-control alongside your code.

Tool-specific sandboxing

A sandbox tuned to each agent and tool — least privilege by default, not one blunt boundary.

Credential protection

Secrets are injected at the boundary and never exposed to the agent. Zeroised on exit.

Fully auditable

Every action lands in a cryptographic, tamper-evident audit trail you can verify.

Member of Technical Staff, Leading Cyber @ OpenAI

正文由 探How 从来源站点 RSS 同步,内容未经改写;遇到排版缺失或需要图片、视频时请以原文为准。

打开原文
去社区讨论 →