AEGIS is a self-hosted, open-source defense layer for LLM apps and agents: prompt injection detection, policy enforcement, and tool/agent permissioning that stops an injected prompt from misusing a tool -- rare among open-source guardrails. What's different: we ran our detectors through adaptive red-team testing (attacks that evolve from prior bypasses) and published the honest result: hardening barely moves the number against an adaptive attacker. Apache 2.0, no signup for the live demo.
Tines The single, secure environment for agents, apps, and automations.
Promoted
Maker
📌
Hey Product Hunt — I built AEGIS, an open-source defense-in-depth platform that sits between your app and any LLM provider: input/output filtering, a policy engine, and (the part I haven't seen elsewhere) tool/agent permissioning with taint tracking, so a prompt injection can't quietly get an agent to misuse a tool.
The part I actually want feedback on: most guardrail projects report a round-1 catch rate against a fixed test set, which tells you almost nothing about a real attacker who adapts. I ran AEGIS's own detectors through adaptive red-team campaigns that breed new attacks from whatever got past the previous round, and published the result even though it isn't flattering: hardened detectors converge to about the same ~48% overall bypass rate as unhardened ones by round 3. Full numbers: github.com/hamidmatiny/aegis/blob/main/RESULTS.md
Try the live demo above, no signup needed — send it a benign question and a prompt injection side by side and watch the difference.
Would genuinely love feedback on: whether the adaptive red-team methodology holds up to scrutiny, what I'm missing on the agent tool-permissioning side, and whether "publish your real bypass rate" is something more guardrail tools should be held to.