Skip to content

AIVF CCS v0.1.0-alpha.3

Pre-release
Pre-release

Choose a tag to compare

View all tags
@wwknow wwknow released this 30 Aug 11:34
v0.1.0-alpha.3
ce5ae64

AIVF CCS v0.1.0-alpha.3

This release adds the first real AI agent runtime integration for AIVF CCS.

AIVF CCS can now sit directly between an OpenAI-compatible tool-calling agent and executable tools, enforcing tool calls through the remote verifier before the tool body is allowed to run.

Highlights

  • Real OpenAI-compatible AI agent integration
  • LLM-generated tool calls routed through AIVF CCS
  • Remote ALLOW / DENY enforcement before tool execution
  • Remote SDK evidence return
  • Independent Ed25519 receipt verification
  • Cross-language Python verifier ↔ Node.js SDK integration
  • Real HTTP tool execution after ALLOW
  • DENY before tool execution for credential-exfiltration patterns
  • RCE detection
  • SSRF / cloud metadata protection
  • Deterministic real-agent integration smoke tests
  • Production-aware OSS/private-runtime scans
  • Improved GitHub onboarding, Quick Start, CI, and architecture documentation

Real Agent Validation

A real OpenAI-compatible tool-calling LLM was validated end-to-end.

ALLOW

LLM-generated http_get tool call
→ AIVF CCS ALLOW
→ tool_executed=true
→ signed receipt
→ signature_valid=true
→ real HTTP response: 200
→ result returned to the LLM

DENY

LLM-generated send_webhook tool call
→ AIVF CCS DENY
→ reason: possible credential exfiltration
→ tool_executed=false
→ signed receipt
→ signature_valid=true
→ denial returned to the LLM

The integration test suite also validates:

RCE request
→ DENY
→ tool_executed=false

Cloud metadata / SSRF request
→ DENY
→ tool_executed=false

Credential-exfiltration request
→ DENY
→ tool_executed=false

SDK

governRemote() now supports:

{
  returnEvidence: true
}

An allowed remote call can return:

output
receipt
verification

This allows agent runtimes to retain the signed CCS evidence associated with the tool authorization decision.

Real Agent Example

See:

examples/real-agent-integration/

It includes:

  • OpenAI-compatible tool calling
  • Remote AIVF CCS enforcement
  • Real HTTP tool execution
  • Safe non-executing shell test path
  • Independent receipt verification
  • RCE, SSRF, and credential-exfiltration test scenarios

Validation

This release passes:

  • 14/14 CCS conformance checks
  • Persistence / replay tests
  • Node.js SDK tests
  • Cross-language Python verifier ↔ Node.js SDK tests
  • Public verifier tests: 5/5
  • Real-agent integration smoke tests
  • Secret scan
  • OSS/public boundary scan
  • Namespace scan

Runtime namespace

issuer:
urn:wwknow:aivf:verifier:prod

audience:
urn:wwknow:aivf:executor:prod

key_id:
aivf-ed25519-1

Live Demo

https://aivf.wwknow.com/

The public demo uses simulated shell execution and never invokes a real shell.

License

Apache License 2.0.

See LICENSE, LICENSE-SCOPE.md, and TRADEMARKS.md.

Assets 2
Loading