Caveat verdict

self-evolving-agent

🟢 Trusted

No high-risk patterns surfaced by the deep scan — automated capability review, not behavioral proof.

Korean-language agent self-improvement skill that analyzes logs and proposes AGENTS.md improvements via a multi-agent pipeline; explicitly states it only proposes changes and never modifies AGENTS.md without user approval, sending reports to Discord.

⚠ Flagged for review — coarse, uncorroborated signal, not a confirmed exploit. Review the config yourself before installing.

Automated static analysis — not a human review. Caveat flags capabilities, not confirmed intent, and can produce false positives. Disagree with this verdict? Use Dispute below.

security

transparency

maintenance

Permission integrity

Installs packages at runtime — transitive dependencies are not auditable

package_install

Findings (18)

Pattern match critical

Pipe to bash — executes piped content as shell commands

SKILL.md · code · | bash

Pattern match critical

Possible hardcoded credential

SKILL.md · code · token: "123456:ABC-DEF...

Pattern match critical

Pipe to python — executes piped content as Python code

CONTRIBUTING.md · code · | python3

Pattern match critical

Recursive delete from root or home — destructive command

docs/demo-video-plan.md · code · rm -rf /

Pattern match critical

Attempts to modify skills or system prompts

docs/clawhub-deployment-checklist.md · code · Edit SKILL.md

Pattern match high

Accesses shell history/config

SKILL.md · code · ~/.zshrc

Pattern match high

Instructs covert action — may act without user awareness

docs/reddit-drafts-v4.md · code · silently

Pattern match high

References sudo — requests elevated privileges

CONTRIBUTING.md · code · sudo

Pattern match high

apt-get install — installs system packages

CONTRIBUTING.md · code · apt-get install

Pattern match high

Accesses system credential store

SECURITY.md · code · keychain

Pattern match medium

References webhook/callback URL

SKILL.md · code · webhook_url

Pattern match medium

References agent memory files

community/README.md · code · MEMORY.md

Pattern match medium

Accesses OpenClaw config/secrets directly

data/fleet/sonnet-analysis.json · prose · downgraded · ~/.openclaw/openclaw.json

Pattern match medium

Popular HTTP library — network access

docs/launch-playbook-v4.md · code · got

Pattern match medium

References agent configuration files

docs/marketing-hn-twitter-v4.md · code · CLAUDE.md

Pattern match low

Python os.environ.get — reads environment variable

scripts/analyze-behavior.sh · prose · downgraded · os.environ.get(

Pattern match low

Python shutil file operation — copies/moves/deletes files

scripts/generate-proposal.sh · prose · downgraded · shutil.move(

Pattern match low

Python urllib.request — network access

scripts/v5/embedding-analyze.sh · prose · downgraded · urllib.request

Why the tier is capped

Execution sink present in raw bytes (Hard Floor: class A/B/D/F). Final tier capped at Caution — cannot be lifted by any downgrade, example-payload opt-in, or allowlist.

Permissions & capabilities

No declared permissions — minimal attack surface.

network_inpackage_install

Is this flag fair?

Check another skill Browse the registry Auditing your own skills or configs? Use the API