Caveat verdict
lobster-continuous-learning-v2
Receives external input AND uses eval
Session-observation learning system that extracts behavioral instincts from Claude Code sessions via hooks and stores them locally; the dynamic_eval capability is for evolving learned patterns into new skills, no exfiltration to external hosts present.
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.
What it does
These are capability combinations: each listed behavior occurs in the skill, but Caveat detects co-occurrence — it does not verify that one flows into another. Read the code to confirm a live chain.
Receives external input AND uses eval — the remote code-injection pattern (data-flow not verified)
LLM01 · LLM05 · ASI01 · ASI05
Permission integrity
agent_memory
Findings (6)
Writes to SKILL.md — self-modifying skill
scripts/instinct-cli.py · prose · downgraded · SKILL.md").write
Accesses sensitive system files
scripts/test_parse_instinct.py · prose · downgraded · /etc/passwd
subprocess execution — runs system commands from Python
scripts/instinct-cli.py · prose · downgraded · subprocess.run(
Instructs covert action — may act without user awareness
scripts/test_parse_instinct.py · prose · downgraded · silently
Python os.environ.get — reads environment variable
hooks/observe.sh · prose · downgraded · os.environ.get(
Python urllib.request — network access
scripts/instinct-cli.py · prose · downgraded · urllib.request
Why the tier is capped
Execution sink present in raw bytes (Hard Floor: class D). 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.
dynamic_evalnetwork_inagent_memory Thanks — recorded.