Security for AI agents that move real money.
AgentGuard evaluates an agent before it deploys, watches every transaction in real time, and pauses it the moment it's exploited — a signed off-chain directive today, and an on-chain broadcast once the pause contract is deployed on Base. The defense the $170k Bankr prompt-injection needed.
DID-based auth · row-level tenant isolation · deterministic, tested core · open protocol
An autonomous agent with wallet access is an attack surface.
Agents on Bankr and Gitlawb sign real transactions from natural-language instructions. A single crafted prompt, a poisoned tool result, or a stale oracle turns that autonomy into a drain — and in May 2026 exactly that cost Bankr roughly $170,000.
What AgentGuard does about it
- →Scores exploitability before deployment.
- →Flags the anomalous transaction as it happens.
- →Fires a signed pause — off-chain directive now, on-chain broadcast once contracts deploy.
- →Turns every caught exploit into protection for all agents.
Evaluate. Monitor. Optimize. Rate. Reward.
Each layer has a real, tested core — and where a data source or contract isn't wired yet, it's labeled, never faked.
Pre-deployment Evaluation
Know an agent's risk before it ever touches a wallet.
A multi-LLM consensus engine generates novel adversarial scenarios tailored to the agent's category, scores how exploitable it is, and signs the report to IPFS.
Real-time Monitoring & Auto-pause
A firewall that halts a compromised agent mid-attack.
Three compute tiers — statistical baselines, a fitted robust anomaly model, and a budget-gated LLM explainer — flag a critical transaction and trigger enforcement: a signed off-chain pause directive, or an on-chain auto-pause once the agent is enrolled and the pause contract is deployed.
LLM Cost Optimization
Cut an agent's inference bill; our fee is a share of the savings.
A strictly tenant-partitioned semantic cache reuses a prior response instead of re-calling the LLM. The fee model only charges on reported savings (per-agent savings metering + settlement are on the roadmap — computed today, not yet billed).
Dynamic Trust Oracle
Public, evidence-aware reputation for every agent.
A live trust score with an honest confidence interval. A brand-new agent reads 50 ± 50 — not a fabricated certainty dressed up as fact.
Vector-driven Bug Bounties
A paid zero-day hardens every agent, instantly.
Researchers submit exploits to an on-chain escrow. On payout the exploit is embedded into a global threat corpus, so the semantic matcher catches it across every agent and chain.
A security product has to be credible.
So it's tested, end-to-end verified against a real database, and it never overstates. Where something isn't real yet, it says so — and refuses to fake it.
A self-sustaining security protocol.
A fair token launch funds the protocol; a creator-fee flywheel and a savings-share fee align every incentive with keeping agents safe. The safer the ecosystem, the more agents deploy — and the more $GUARD accrues.
Deploy the guarded way.
Evaluate your agent, wire up real-time monitoring, and let AgentGuard defend it from transaction #1.