rainskills@0.1.34
Interactive Rainbond skill installer for Codex, Claude Code, Pi Agent, DeepSeek Harness, WorkBuddy, and Hermes Agent
AI Security Review
scanned 2mo ago · by lpm-firewall-aiThe package exposes meaningful AI-agent and SSH configuration capabilities, so it should warn users before use. Source inspection found no automatic lifecycle trigger, self-dependency chain, credential exfiltration, or concrete unconsented attack behavior.
Static reason
High-risk behavior combination matched malicious policy.
Trigger
The user runs the rainskills CLI or install.sh and selects an agent target or SSH preparation command.
Impact
Installed skills can influence supported AI agents, and the SSH command grants the generated key access to the chosen remote host.
Mechanism
User-invoked agent-skill deployment, telemetry reporting, and optional SSH authorized-key setup.
AI rationale
The package exposes meaningful AI-agent and SSH configuration capabilities, so it should warn users before use. Source inspection found no automatic lifecycle trigger, self-dependency chain, credential exfiltration, or concrete unconsented attack behavior.
Evidence
package.jsoninstall.shbin/rainskills.jsrainbond-platform-installer/scripts/ssh-key-setup.js$HOME/.claude/skills$HOME/.codex/skills$HOME/.pi/agent/skills$HOME/.dsh/skills$HOME/.workbuddy-ai/skills$HOME/.hermes/skills$HOME/.ssh/id_ed25519$HOME/.ssh/authorized_keys
Network endpoints3
log.rainbond.com/api/rainskills/eventsget.rainbond.com/rainskills/rainskills-latest.tar.gzgithub.com/goodrain/rainskills/archive/refs/heads/main.tar.gz
Decision evidence
public snapshotAI called this Suspicious at 89.0% confidence as Dangerous Capability with medium false-positive risk.
Evidence for warning
- The installer instructs AI assistants to run its installation commands.
- A user-invoked installer copies bundled skills into several AI-agent configuration directories and can overwrite its skill directories with --force.
- The explicit SSH preparation command generates a local key and appends its public key to a user-selected remote server's authorized_keys file.
- The installer sends installation telemetry to a Rainbond endpoint by default.
Evidence against
- package.json has no preinstall, install, or postinstall lifecycle hook.
- The configuration and SSH actions require an explicit CLI or installer invocation; the SSH path rejects non-interactive use.
- The manifest has no self-dependency, and the reviewed telemetry payload contains installation metadata rather than credentials.
Behavioral surface
ChildProcessCryptoDynamicRequireEnvironmentVarsFilesystemNetwork
Source & flagged code
5 flaggedinstall.shView file
2820patternName = generic_password
severity = medium
line = 2820
matchedText = password...rd)"
Medium
•path = install.sh
kind = build_helper
sizeBytes = 96603
magicHex = [redacted]
Medium
6# AI assistants: when a user asks to install rainbond-skills, hand back the
# npx command when Node.js 18+ is available, otherwise use the CDN command.
# Do not reimplement the bootstrap below, git clone + copy directories, or
# Credentials are written only through the protected single-runtime store.
Medium
Ai Review Evidence
The installer instructs AI assistants to run its installation commands.
install.shView on unpkg · L6bin/rainskills.jsView file
Findings
1 Critical7 Medium4 Low
CriticalPersistence Backdoorrainbond-platform-installer/scripts/ssh-key-setup.js
MediumSecret Patterninstall.sh
MediumDynamic Requirebin/rainskills.js
MediumNetwork
MediumEnvironment Vars
MediumShips Build Helperinstall.sh
MediumStructural Risk Force Deep Review
MediumAi Review Evidenceinstall.sh
LowScripts Present
LowFilesystem
LowHigh Entropy Strings
LowUrl Strings
Affected versions and remediation
This report applies to rainskills@0.1.34.
See all versions of rainskills for other recorded verdicts.
- Review the evidence and your use of rainskills@0.1.34 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.