Hazmat gives AI coding agents full autonomy by containing them at the OS level — dedicated macOS user, kernel sandbox (Seatbelt), pf firewall, DNS blocklist, supply chain hardening (npm ignore-scripts), and automatic snapshots.
I built it because manual approval mode is the worst of both worlds: fake security (any prompt injection bypasses it) and a productivity killer (you become the bottleneck). With OS-level containment in place, --dangerously-skip-permissions becomes safe by default.
Setup/rollback ordering is formally verified with TLA+ (26,905 states, 3 real bugs found). Works with Claude Code, OpenCode, Codex, or any terminal agent.
I built it because manual approval mode is the worst of both worlds: fake security (any prompt injection bypasses it) and a productivity killer (you become the bottleneck). With OS-level containment in place, --dangerously-skip-permissions becomes safe by default.
Setup/rollback ordering is formally verified with TLA+ (26,905 states, 3 real bugs found). Works with Claude Code, OpenCode, Codex, or any terminal agent.
brew install dredozubov/tap/hazmat
Full write-up: https://codeofchange.io/how-i-made-dangerously-skip-permissi...