Repository navigation
Releases
Releases · invarcore/canary-fabric
v0.2.1: Multi-Agent Blast Radius Breaker (ASI08) & Strike Decay
Canary Fabric v0.2.1
Key Capabilities
- OWASP ASI08 Multi-Agent Blast Radius Circuit Breaker: Implements BlastRadiusCircuitBreaker with sub-millisecond rolling FIFO window tool-call velocity monitoring per agent.
- Progressive Strike Decay: Automatically decays accumulated strikes over configurable quiet windows, preventing permanent lockouts from transient traffic spikes.
- Half-Open Probationary Recovery: Safely transitions tripped agents through a probationary period before fully restoring active status.
- Bounded Ephemeral Agent Pruning: Automatically prunes idle and dead agents beyond capacity limits, guaranteeing deterministic O(1) memory consumption in high-churn multi-agent swarms.
- 100% Unit Test Coverage: Verified across 100 unit tests with sub-millisecond tripwire latency.
Full Changelog: v0.2.0...v0.2.1
Canary Fabric v0.2.0: Invisible Cryptographic Tripwires & Streaming Circuit Breakers
Canary Fabric v0.2.0: Invisible Cryptographic Tripwires & Streaming Circuit Breakers for RAG & Autonomous AI Agents
canary-fabric is a deterministic, wire-level data exfiltration defense system for Retrieval-Augmented Generation (RAG) and autonomous AI agent runtimes.
⚡ 60-Second Quickstart (Zero External Services)
Run the self-contained verification demo in <1 second with zero infrastructure dependencies:
git clone https://github.com/invarcore/canary-fabric.git && cd canary-fabric
pip install -e .
python examples/quickstart.py🛡️ Core Capabilities & Invariants
- 4-ary Zero-Width Cryptographic Steganography:
- Injects 64-bit HMAC-derived canary tokens (
\u200B,\u200C,\u200D,\uFEFF) invisibly into retrieved chunks. - 100% imperceptible to human reviewers; preserved across standard LLM tokenizers.
- Injects 64-bit HMAC-derived canary tokens (
- Sub-Millisecond Wire-Level Streaming Circuit Breaker:
SlidingWindowStreamBufferscans outbound Server-Sent Events (SSE) and agent tool-call JSON in real-time.- Micro-benchmark: 17.73 µs scan latency per chunk (< 0.05 ms Time-to-First-Token tax, 50,000x faster than synchronous LLM/NER guardrails).
- Severs rogue egress streams before customer PII or confidential IP is rendered or transmitted.
- Synthetic Decoy Honeytokens:
- High-fidelity synthetic credentials (
sk_live_canary_..., PostgreSQL connection strings, AWS secrets, canary emails) for honeypot documents and data trap monitoring.
- High-fidelity synthetic credentials (
- Tamper-Evident Forensic Leak Certificates:
- Emits HMAC-SHA256 authenticated forensic certificates cryptographically binding each breach to its source document chunk, tenant namespace, and prompt hash.
- Turnkey Ecosystem & Framework Adapters:
- Native drop-in adapters for LangChain, LlamaIndex, FastMCP, Knowledge Fabric, Intent Fabric, and OpenAI/LiteLLM streaming wrappers.
🧪 Quality Gates & Verification Suite
- Unit & Integration Suite: 90/90 passed (100% green across Python 3.11, 3.12, 3.13).
- Code Coverage: 97.41% (exceeding strict 90% threshold).
- Docker Verification: Hermetic Docker Compose test suite verified.
- Packaging: Built and signed wheel and sdist distributions attached.
Full Changelog: https://github.com/invarcore/canary-fabric/commits/v0.2.0