Trust & Security — AgentLeash

Core Security Principles

These aren't aspirational. They're enforced in code, verified on-chain, and audited on every action.

🚫

Fail Closed

Any error during verification results in denial. Network timeout? Deny. Cache miss? Verify, then deny if unresolvable. Ambiguity is always denial.

⛓️

Dual Verification

Permissions live both on-chain (immutable) and in our database (fast). Both must agree. If they don't, the agent is blocked until resolved.

🔥

Instant Revocation

Burn a token on-chain and the agent is immediately inert. No propagation delay, no grace period. One transaction, zero access.

🎭

Privacy by Default

Real agent names never appear on the public blockchain. Decoy identifiers are generated automatically. Only you can link them back.

🔐

Zero-Knowledge Credentials

Agents never see raw API keys or passwords. The vault injects credentials at the moment of use, in-memory only, then discards them.

📋

Complete Audit Trail

Every tool call, every decision, every token lifecycle event is logged. Exportable, searchable, and tamper-evident via blockchain anchoring.

How Verification Works

Every single tool call goes through this pipeline. No exceptions.

🤖

Agent Tool Call

read, exec, message...

→

🔌

Runtime Plugin

Intercepts before execution

→

⚡

Verify API

Cache → DB → Chain

→

✅

Allow

Tool executes normally

🚫

Deny

Blocked with reason

Security Features

Defense in depth. Every layer adds protection.

🔗

Blockchain-Backed Permissions

Permissions are minted as NFTs on Base (ERC-721) or Algorand (ASA). Immutable, verifiable, and publicly auditable.

🔒

Encrypted Permission Tokens

AES-256-GCM encryption with per-token key derivation (HKDF). Even if the database is compromised, permissions remain encrypted.

🔑

Session Binding

Tokens can be bound to a specific session. A stolen token ID is useless in a different session context.

🚫

Non-Transferable Tokens

PATs cannot be transferred between wallets. Only the issuer can move them via revocation. No token marketplace, no unauthorized sharing.

⏰

Auto-Expiration

Tokens auto-expire after a configurable period (default: 1 year). The chain-sync service checks every 5 minutes.

🛡️

Data Protection Blocks

Block agents from accessing passwords, credit cards, PII, or health data at the content level, not just the tool level.

🌍

GeoFence + OFAC Screening

14 OFAC-sanctioned countries blocked at signup. OFAC SDN list screened (9,180 entries, weekly refresh). MaxMind GeoIP2 for production.

🚦

Rate Limiting

Redis sliding-window rate limits on every endpoint. Per-IP, per-user, per-agent. 429 with Retry-After. Fail-open if Redis is down (traffic still flows).

Compliance Ready

Built for organizations that answer to regulators, not just users.

🏛️

OFAC Compliant

GeoFence + SDN screening at signup. 14 sanctioned countries blocked.

📋

Full Audit Trail

Every verification logged. CSV/JSON export. Blockchain-anchored timestamps.

🔐

Encryption at Rest

AES-256-GCM for permissions. AWS KMS for key management. HKDF derivation.

🏥

Data Protection

Content-level PII, HIPAA, PCI blocks. Agent can't see what you don't allow.

<5ms

Cached verification latency

Blockchains supported

Interceptable tools

100%

Tool calls verified

Security FAQ

What happens if the verification API goes down?▾

The agent is blocked. AgentLeash fails closed by design. No verification = no tool execution. This is configurable (you can set fallback to "allow" for non-critical agents), but the default is always deny.

Can an agent bypass the permission check?▾

No. The AgentLeash plugin runs at the kernel level of the agent runtime (priority 100, before all other plugins). Every tool call passes through it. There is no code path that skips verification when the plugin is loaded.

What data is stored on the public blockchain?▾

By default, only a pseudonymous identifier (SHA-256 hash) and a permissions hash. Real agent names, full permission schemas, and all sensitive data stay in our encrypted database. You can opt in to showing the real name on-chain, but it's off by default.

How fast is revocation?▾

Immediate. When you revoke a token, the on-chain transaction processes in 2-4 seconds. The Redis cache is invalidated instantly. The agent's next tool call will fail verification. There is no grace period or propagation delay.

Does the agent ever see my API keys or passwords?▾

Never. The Secure Vault stores credentials encrypted with AES-256-GCM. When an agent needs to call an API, the runtime injects the credential at the moment of the HTTP request, in-memory only. The agent sees the API response, never the key.

Is AgentLeash suitable for government use?▾

Yes. AgentLeash was designed with government and enterprise compliance in mind. OFAC screening, GeoFence enforcement, HIPAA-level data protection blocks, full audit trails, and blockchain-verified immutable records. We're actively engaged with federal agencies on pilot programs.

Where is data stored?▾

AWS US-East-2 (Ohio). RDS PostgreSQL with Multi-AZ deployment. ElastiCache Redis for caching. All data encrypted in transit (TLS) and at rest (AES-256). On-chain data lives on Ethereum L2 - Base (Ethereum L2) and/or Algorand.

Built for Zero Trust