negotium@0.14.3
Install the Negotium multi-agent runtime and CLI with one package
AI Security Review
scanned 1mo ago · by lpm-firewall-aiThe code shows a real automatic native-payload installation path and high-capability agent integration, so it should not be marked clean. Static inspection found no concrete credential exfiltration, destructive action, self-dependency chain, or broad AI-agent control-surface mutation sufficient for a block.
Decision evidence
public snapshot- The install lifecycle automatically runs two downloaders.
- The browser downloader fetches a native executable, verifies a fixed hash, and marks it executable under the user's Negotium state directory.
- Runtime code can create a temporary Codex wrapper that invokes Codex with hook trust bypassed.
- The bundled MCP factory exposes local MCP endpoints; the large agent runtime has substantial command and file-access capability.
- The CLI starts the node only after a user invokes its serve command.
- No self-dependency is declared.
- No source evidence shows install-time credential harvesting, external secret exfiltration, or mutation of Codex or Claude configuration.
- The observed MCP URLs are loopback endpoints, not an unofficial credential gateway.
- The downloaded binaries have release-specific SHA-256 checks before installation.
Source & flagged code
9 flagged · loading sourcePackage defines install-time lifecycle scripts.
package.jsonView on unpkgInstall-time lifecycle script is not statically allowlisted and needs review.
package.jsonView on unpkgPackage source references child process execution.
dist/hosted-agent.jsView on unpkg · L9Findings
Affected versions and remediation
This report applies to negotium@0.14.3.
See all versions of negotium for other recorded verdicts.
- Review the evidence and your use of negotium@0.14.3 before allowing it. Restrict the permissions described in this report.
- Choose an independently verified alternative or release. This report does not establish that other versions are safe.
Evidence last updated: .
Think this verdict is wrong? Dispute it on GitHub or read how reports are produced.