Reef Prompt Guard

Detect and filter prompt injection attacks in untrusted input. Use when processing external content (emails, web scrapes, API inputs, Discord messages, sub-agent outputs) or when building systems that accept user-provided text that will be passed to an LLM. Covers direct injection, jailbreaks, data exfiltration, privilege escalation, and context manipulation.

This skill appears to be what it claims: a regex-based prompt-injection filter implemented as a local Python script. Before installing/use: 1) Prefer invoking the filter module directly (import/call) rather than interpolating untrusted text into shell commands — the Node.js execSync example in SKILL.md can be unsafe and lead to command injection if input isn't properly escaped. 2) Understand the tool's limitation: it's regex-based and will miss novel/semantic attacks; consider adding a classifier or anomaly/perplexity checks for ambiguous inputs. 3) Review and test the pattern lists in scripts/filter.py and references/attack-patterns.md to ensure no false-positives block legitimate content and to tune context multipliers. 4) Keep the script on a secure path and avoid running it with elevated privileges. If you need stronger guarantees (e.g., in production-facing pipelines or multi-agent systems), perform adversarial testing and consider layered defenses (sandboxed processing, dual-LLM architecture, strict escaping when calling subprocesses).

Type: OpenClaw Skill Name: reef-prompt-guard Version: 1.0.0 This skill bundle, 'reef-prompt-guard', is a security tool designed to detect and filter prompt injection attacks. The `scripts/filter.py` script uses regular expressions to identify patterns associated with various prompt injection techniques (e.g., instruction override, data exfiltration, command execution attempts) and then sanitizes the input. The `SKILL.md` and `references/attack-patterns.md` files serve as documentation, explaining the purpose, usage, and underlying attack patterns this skill defends against. There is no evidence of malicious intent, data exfiltration, unauthorized execution, or prompt injection against the OpenClaw agent itself; rather, the skill actively works to prevent these types of attacks.