harness

No clear malware payload or direct data exfiltration is present. The script is an automation harness, but it presents a significant security risk because it sends untrusted repository and temporary-file content to Claude and invokes both Worker and Reviewer with --permission-mode bypassPermissions. A compromised document or context file could prompt the agent to perform unauthorized actions. Use restricted permissions, validate prompts and changes, protect the temporary context file, and retain complete logs.

The script appears intended as an AI-assisted repository inspection harness, not as direct malware. Its main security issue is high-impact prompt injection and unauthorized modification risk: repository documents and /tmp/harness-context.txt are inserted verbatim into an agent prompt, while Claude is run with bypassPermissions. Restrict the agent to a narrowly scoped workspace or explicit approval, treat all document content as untrusted, secure the temporary context file, and avoid forwarding secrets or sensitive files.