Skip to content
WeChat · Goumang Security Lab·Indexed by this site · Originally published· Tools →· Its place in the scan record →

Jumang hands-on test: SolonGate intercepts coding agent tool calls at the hook layer

「AI Agent安全+工具调用拦截+开源实测」SolonGate:编程Agent的执行前闸门

AI summary

Jumang Security Lab hands-on tested the open-source project SolonGate: a local security gateway written in Go that intercepts coding agent tool calls before execution through three hooks, PreToolUse, PostToolUse, and Stop, and supports Claude Code, Codex CLI, Antigravity, and OpenCode. Following the actual order of the code, the author laid out nine decision checks, covering local policy, self-protection, sandbox requirements, outbound network checks, parameter DLP, rate limiting, and OPA policy evaluation; the sandbox is based on Landlock, Seatbelt, and icacls on Linux, macOS, and Windows respectively. The article summarizes public issues opened by the maintainers, such as #104 signature verification fail-open, #193 quoted paths containing spaces being split and bypassing path deny rules, and #220 icacls failing on Windows yet still running, and, drawing on the semantic divergence between the two evaluators in #261, emphasizes that fallback paths must have the same semantics as the main path.

Read originalmp.weixin.qq.com