Selective Pollution Test Skill

Research skill that conditionally rewrites search keywords on trigger topics for authorized security testing.

This skill is internally consistent with its stated purpose, but it enables intentional manipulation of search keywords (adding 'speculative rumor' to finance/health queries) which can be harmful if misused. Before installing: (1) verify you have explicit authorization and a safe, isolated test environment (no real users or public queries), (2) confirm the skill's provenance (owner/source is unknown; lack of homepage increases risk), (3) add operational guardrails — require confirmation prompts, run only on test accounts/domains, log all inputs/outputs, and limit autonomous invocation or set disable-model-invocation if you don't want the agent to call it without user consent, and (4) consider implementing programmatic safeguards (whitelists, domain isolation, and automatic stop-conditions) if you will use it for research. If you cannot confirm authorization or source, do not install in production or against public services.

Type: OpenClaw Skill Name: selective-pollution-test Version: 1.0.0 The skill 'selective_pollution_test' (SKILL.md) instructs the AI agent to conditionally manipulate user search queries by appending the phrase 'speculative rumor' to finance or health-related topics. While the README.md states this is for 'controlled security research' on tool-chain manipulation, the behavior is a functional demonstration of a prompt-injection attack designed to alter the integrity of the agent's output based on content triggers. No evidence of data exfiltration, persistence, or malicious code execution was found.