A plumber-turned-coder has built Pipelock, a security harness for AI coding agents that sits between your AI assistant and the outside world. The single Go binary prevents credential exfiltration by scanning all outbound traffic for secrets, blocking unauthorized network destinations, and monitoring workspace files for unauthorized changes.

Traditional static scanners check code before installation but can't catch a trusted tool that decides to phone home at runtime. Pipelock addresses this gap by implementing capability separation - the agent process with secrets is network-restricted while a separate fetch proxy handles web browsing. Every request runs through a 7-layer scanner pipeline including DLP patterns, entropy analysis, and domain blocklists.

Recent research found that 283 out of 3,984 published agent skills (about 7%) were leaking credentials. The tool caught a skill embedding a Slack token in debug logs heading to an external endpoint - something that would have gone unnoticed without runtime scanning. Pipelock offers three security modes from airtight restriction to audit logging, and includes OWASP Agentic Top 10 coverage for threats like prompt injection and privilege escalation.