Arc Shield

v1.0.0

Output sanitization for agent responses - prevents accidental secret leaks

0· 935·0 current·0 all-time
MIT-0
Download zip
LicenseMIT-0 · Free to use, modify, and redistribute. No attribution required.
Security Scan
VirusTotalVirusTotal
Benign
View report →
OpenClawOpenClaw
Benign
high confidence
Purpose & Capability
The skill's name, description, and included scripts (Bash + Python) align with an output-sanitization purpose. Requested runtime (bash, python3) is appropriate. One minor inconsistency: documentation refers to config/patterns.conf as the pattern database but that file is not present in the provided file manifest — the scripts default to loading ../config/patterns.conf, so installation will need that file (docs claim it exists).
Instruction Scope
SKILL.md and examples limit actions to scanning/sanitizing outbound messages, running locally, and integrating as a pre-send hook or wrapper. The instructions do not tell the agent to read unrelated system files or to transmit data to external endpoints. Integration examples do append a local log entry (~/.openclaw/logs/arc-shield-blocks.log) when blocking — this is reasonable and documented.
Install Mechanism
There is no automated install spec (instruction-only install via git clone / manual copy), which is low-risk. The scripts themselves make no external network calls. Note: the documentation and code expect a config/patterns.conf file; that file is referenced but not present in the provided manifest — you'll need to ensure that config is supplied when installing.
Credentials
The skill requests no environment variables or credentials. It analyzes text passed to it rather than reading secrets from the environment. It does not require unrelated credentials, so the requested environment access is proportionate.
Persistence & Privilege
The skill is not force-included (always:false) and does not modify other skills or system components automatically. Integration requires the operator to add a pre-send hook or wrapper, which is a deliberate, user-controlled action. Example hooks log blocked attempts to a local path — this is described in the docs.
Assessment
Arc Shield appears to be a coherent, local output-sanitizer that uses regex and entropy heuristics. Before installing: 1) Verify the config/patterns.conf file is present (docs reference it but it wasn't in the manifest); without it detection may be incomplete. 2) Review scripts (scripts/arc-shield.sh and scripts/output-guard.py) yourself — they operate locally and do not call external network endpoints, but you should confirm there are no edits that add remote posting. 3) Run the included tests (./tests/quick-test.sh) in a safe environment to validate behavior and tune patterns/entropy threshold to avoid false positives. 4) When integrating as a pre-send hook, ensure internal channels are excluded (examples show this) and check where blocked attempts are logged (e.g., ~/.openclaw/logs/arc-shield-blocks.log) to avoid leaking sensitive metadata. 5) Don’t rely on this alone — continue to train agents to avoid emitting secrets and treat redaction as a safety net. If you want higher assurance, request the missing config file and full, un-truncated copies of the scripts so you can audit the complete code paths.

Like a lobster shell, security has layers — review code before you run it.

latestvk97dnkcpd5bwmg7gbkjrt8jqgx80t2ms

License

MIT-0
Free to use, modify, and redistribute. No attribution required.

Comments