Infrastructure Release Notes v10.1.5

The v10.1.5 release introduces 5 new infrastructure modules that complete the ABS Core governance pipeline for MCP tool calls.

1. MCP Bridge - Tool Call Interception

The MCP Bridge intercepts Model Context Protocol (MCP) tool calls before execution, applying the same governance pipeline as the LLM Proxy.

Endpoints

Evaluation Pipeline

Tool Call -> DSL Policy Engine -> Regex Policy Engine -> Entropy Analysis -> Bond/Slash -> WAL Log -> Verdict

Usage Example

curl -X POST https://api.abscore.app/v1/mcp/tool-call \
  -H "Content-Type: application/json" \
  -d '{"name":"fs.write","arguments":{"path":"/etc/passwd","content":"test"}}'

# Response (403):
# {"verdict":{"decision":"DENY","policyResult":{"status":"DENIED","reason":"Access to system-critical paths is forbidden"}}}

Invariants

• MCP-I1: Every tool call is evaluated before execution
• MCP-I2: DENIED calls never reach the tool server
• MCP-I3: All decisions are logged in the WAL (hash-chain)
• MCP-I4: Entropy above threshold triggers automatic FLAG

2. Declarative Policy Language (DSL)

Replaces hardcoded regex with declarative JSON policies, similar to OPA/Rego but optimized for AI.

Format

{
  "version": "1.0",
  "policies": [
    {
      "name": "block_credential_access",
      "match": {
        "tools": ["fs.write", "fs.delete"],
        "args_contain": [".env", ".pem", ".key"]
      },
      "action": "DENY",
      "risk": 1.0,
      "reason": "Modification of credential files is forbidden"
    }
  ]
}

SYSTEM_POLICIES (built-in, v10.1.5)

Invariants

• DSL-I1: First DENY wins (evaluation in order)
• DSL-I2: No match = ALLOW (default open)
• DSL-I3: Policies are stateless
• DSL-I4: Definition errors are rejected at parse time, not runtime

3. Agent Identity (JWT)

Cryptographic authentication for AI agents via JWT HMAC-SHA256, with claims describing permissions and risk budget.

Claims

API

import { generateAgentToken, verifyAgentToken, isToolAllowed } from './security/agent-identity';

// Generate token (1h TTL)
const token = await generateAgentToken({
  agentId: 'agent-abc',
  tenantId: 'tenant-123',
  tools: ['fs.read', 'llm.*'],
  riskBudget: 50,
  bonded: true,
}, process.env.ABS_SECRET_KEY);

// Verify (sub-1ms, no external calls)
const { valid, claims } = await verifyAgentToken(token, secret);

// Check tool permission
const allowed = isToolAllowed(claims, 'fs.write'); // false

Invariants

• AID-I1: HMAC-SHA256 (Web Crypto API, edge-compatible)
• AID-I2: Expired tokens are always rejected
• AID-I3: Claims are immutable for the token lifetime
• AID-I4: Verification sub-1ms (no external calls)

4. Proxy Rate Limiter

Native per-agent rate limiting with sliding window in-memory.

Default Limits

HTTP Headers

X-RateLimit-Limit: 100
X-RateLimit-Remaining: 87
X-RateLimit-Reset: 1708266790
Retry-After: 42  (only on 429)

Key Extraction (priority)

1. Agent JWT sub claim
2. API Key (first 16 chars)
3. CF-Connecting-IP / X-Forwarded-For

Invariants

• RL-I1: Rate limits are enforced before any processing
• RL-I2: Memory is bounded (max 10k tracked identities)
• RL-I3: Expired entries are garbage collected periodically

5. Python SDK

Official SDK for Python 3.10+ with zero dependencies (stdlib only).

Installation

pip install abs-sdk

Quick Start

from abs_sdk import ABSClient

client = ABSClient(
    dsn="abs://my-workspace.abscore.app",
    token="abs_pat_xxx"
)

# Evaluate tool call
verdict = client.evaluate_tool("fs.write", {"path": "/etc/passwd"})
if verdict.decision == "DENY":
    raise PermissionError(f"Blocked: {verdict.policy_result.reason}")

# Heartbeat
client.heartbeat("agent-123")

# Bond
bond = client.get_bond("agent-123")
print(f"Balance: ${bond.balance}")

Features

• evaluate_tool() / evaluate_batch() - MCP evaluation
• heartbeat() / get_agent() / register_agent() - agent management
• create_bond() / get_bond() / release_bond() - bond/slash
• audit_export() - audit trail export
• calculate_entropy() - local Shannon entropy (offline)

Evaluation Architecture v10.1.5

Each tool call passes through the following layers in sequence:

1. MCP Bridge - receives POST /v1/mcp/tool-call
2. DSL Policy Engine (primary) - evaluates SYSTEM_POLICIES (7 rules) + tenant policies
3. Regex Policy Engine (secondary) - financial and complementary rules
4. Entropy Analysis - Shannon entropy above 4.5 triggers automatic FLAG
5. Bond/Slash - on DENY, slashes the agent bond
6. WAL Audit Log - all decisions recorded in immutable hash-chain

All modules are live in production at api.abscore.app since v10.1.5.