Caveat verdict
subconscious
A bounded self-improvement bias layer that reads local learnings files and promotes typed mutations to an agent behavioral store; core identity is stated as immutable without manual override and it is scoped to a named user agent.
⚠ 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.
Findings (6)
Recursive delete from root or home — destructive command
references/INSTALL.md · code · rm -rf ~
References agent configuration files
README_zh.md · code · CLAUDE.md
Instructs covert action — may act without user awareness
README.md · prose · downgraded · quietly
References agent memory files
self-improving-agent/references/openclaw-integration.md · code · MEMORY.md
Popular HTTP library — network access
subconscious/schema.py · prose · downgraded · got
Python os.environ.get — reads environment variable
subconscious/store.py · prose · downgraded · os.environ.get(
Why the tier is capped
Execution sink present in raw bytes (Hard Floor: class B). 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.
Is this flag fair?
Thanks — recorded.