Menu⌄

Security first

Safety Radar beta now; performance benchmarks, private registries, and runtime skill recommendations next.

Alert preview · safety before automation

See the alert before an agent skill quietly changes.

This is the private watchlist experience: SkillRadar turns source diffs into human-review alerts when a watched skill adds commands, credentials, domains, risk flags, or other safety-sensitive behavior.

Get my agent stack watched Browse ranked AI skills

SkillRadar Alert

Agent Browser

High review

New credential requirement detected

Browser automation skill changed before your agent ran it.

New env var

AGENT_BROWSER_ENCRYPTION_KEY

New command

npm install -g agent-browser

Why it matters

Browser automation can touch logged-in sessions, private pages, screenshots, forms, cookies, and local browser state.

Recommended action

Review the diff and approve manually before letting an agent install, update, or execute this skill.

sample watchers

matched alert items

2026-05-04

source refresh

Sanitized digest sample

What a beta user receives

Real alerts are generated from the user’s watchlist. This sample uses sanitized watcher names and public source links so users can understand the safety workflow before joining.

Join watchlist beta →

Watcher: agent-security-team

2 matched changes · private digest destination hidden in production

human review

github.com/openclaw/openclaw@main

High risk

openclaw-qa-testing

New shell command

pnpm test:qa -f package_label=openclaw@YYYY.M.D-beta.N -f package_spec=openclaw@YYYY.M.D-beta.N -f provider_mode=mock-openai

Source file ↗Skill report →

github.com/openclaw/openclaw@main

High risk

openclaw-parallels-smoke

New test command

gtimeout --foreground 150m pnpm test:parallels:npm-update -- --beta-validation beta4 --model openai/gpt-5.4 --json

Source file ↗Skill report →

Product framing

Security-style monitoring for agent skills.

SkillRadar is not claiming every alert is malicious. It is the early-warning layer: when third-party agent instructions quietly add a command, credential, domain, remote installer, or higher-risk behavior, the human sees it before automation does.