I built a deterministic
Most LLM guardrails (like Llama Guard or NeMo) add 500ms–1500ms of latency per call. For real-time autonomous AI agents, that delay kills the user experience. I built CanaryGuard: a deterministic heuristic firewall that
Most LLM guardrails (like Llama Guard or NeMo) add 500ms–1500ms of latency per call. For real-time autonomous AI agents, that delay kills the user experience.
I built CanaryGuard: a deterministic heuristic firewall that inspects inputs in <1ms line-rate before passing them to the model context.
What it Catches
- Chat-template delimiter smuggling (
[INST],<|im_start|>,<<SYS>>) - Concealed unicode obfuscation (zero-width spaces
\u200B, bidi overrides) - Hidden HTML/DOM injection (
display:none, opacity tricks) - Asset exfiltration & wallet drain commands
Open Red-Team Challenge
To test how far deterministic rules hold up against real jailbreaks, I seeded an on-chain $5.00+ USDC bounty vault on Base & Solana. If your payload slips past strict mode with status clean, you win the vault.
👉 Test your payloads in the live simulator:
https://canaryguard-402.onrender.com/bounty.html
(Now includes a Sanitizer Mode toggle to see before/after text scrubbing live).
Curious to hear what injection payloads you can use to break it!
Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes — full credit and traffic to the original publisher.