MARKETS RISK-OFF — stocks red, yields ripping, bonds selling off. This is exactly when a hijacked AI agent panic-dumps your book or wires shares out. Watch SENTRY halt it ▸
◇ RUNTIME AI SECURITY · A CIRCUIT BREAKER FOR TRADING AGENTS

An antivirus for AI agents that trade.

SENTRY LAYER is a runtime security layer for autonomous agents trading tokenized stocks on Robinhood Chain — WAF · EDR · DLP for the agent stack. It inspects every prompt, market feed, model output and on-chain trade inline, and blocks the ones that shouldn't execute.

OWASP LLM Top 10 MITRE ATLAS Robinhood Chain
INPUT GUARDinjected orders · jailbreaks
OUTPUT · DLPportfolio · positions · PII
TRADE FIREWALLtaint-tracked trades & transfers
MARKET STRESS · RISK-OFF · LIVE
SPY ▼ -2.9% NVDA ▼ -4.1% TSLA ▼ -6.2% AAPL ▼ -3.1% MSFT ▼ -2.4% US 10Y ▲ +19bp US 30Y ▲ +23bp TLT ▼ -3.4% VIX ▲ +31% SPY ▼ -2.9% NVDA ▼ -4.1% TSLA ▼ -6.2% AAPL ▼ -3.1% MSFT ▼ -2.4% US 10Y ▲ +19bp US 30Y ▲ +23bp TLT ▼ -3.4% VIX ▲ +31%

Antivirus was built for files.
Attacks now travel in meaning.

A classic AV matches deterministic bytes. Attacks on a trading agent are semantic, probabilistic and contextual — the same message is harmless or catastrophic depending on whether the agent can move your tokenized shares. You can't sign a payload that rephrases itself infinitely.

So we keep what carries over — layered defense, hybrid detection, a live threat feed, quarantine, telemetry — and rebuild the rest as runtime security for on-chain agents.

Classic AV

  • file signatures
  • known-bad bytes
  • static scan
  • on-disk quarantine

SENTRY LAYER

  • semantic detection
  • intent & provenance
  • inline runtime
  • trade-level firewall

The threat model

Eleven classes across input, market data, output, and on-chain trades — mapped to OWASP LLM Top 10 & MITRE ATLAS.

Cheap filters first. Expensive models last.

A hybrid conveyor — each stage only escalates what the previous couldn't resolve.

01

Normalize

base64 · hex · zero-width · NFKC · homoglyph

02

Signatures

regex + fuzzy TI feed

03

Embeddings

similarity to attack corpus

04

Guard model

fine-tuned classifier

05

LLM-as-judge

escalation for the grey zone

Try the working console.

The whole detection pipeline runs client-side in your browser. No data leaves the page.

Launch App ▸