As cited
Copy frozen at (site build).
threat intel
Living off the coding agent: Two tales of tunnels and LaunchAgents
Elastic Security researchers documented a case where a coding agent (Claude Code) on a macOS developer endpoint was used to establish reverse tunnels, expose local applications to the internet, and install LaunchAgent persistence mechanisms. The activity occurred across multiple days in July 2026 and combined legitimate dual-use tools (Cloudflare Quick Tunnels, localhost.run, ngrok) with high-severity outcomes including credential exposure and persistence installation, making detection and triage challenging when trusted vendor-signed tools serve as the process parent.
Why it matters: Detection engineers and security analysts on macOS endpoints need to recognize that agent-parented reverse tunnels, credential exfiltration, and LaunchAgent persistence represent high-severity compromise signals even when the parent process is a trusted coding agent, since dual-use tools can enable unauthorized remote access to local services.
- Source published
- First seen by Cybersecurity Tracker
Source attribution
Correction
Correction recorded as of .
threat intel
Living off the coding agent: Two tales of tunnels and LaunchAgents
Elastic Security researchers documented a case where a coding agent (Claude Code) on a macOS developer endpoint was used to establish reverse tunnels, expose local applications to the internet, and install LaunchAgent persistence mechanisms. The activity occurred across multiple days in July 2026 and combined legitimate dual-use tools (Cloudflare Quick Tunnels, localhost.run, ngrok) with high-severity outcomes including credential exposure and persistence installation, making detection and triage challenging when trusted vendor-signed tools serve as the process parent.
Why it matters: Detection engineers and security analysts on macOS endpoints need to recognize that agent-parented reverse tunnels, credential exfiltration, and LaunchAgent persistence represent high-severity compromise signals even when the parent process is a trusted coding agent, since dual-use tools can enable unauthorized remote access to local services.
- Source published
- First seen by Cybersecurity Tracker