Bartholomew — Interactive Trajectory & Kill-Switch Inspector

1. Threat Severity Classification Taxonomy CRITICAL EXPLOIT PROTECTION

Bartholomew classifies every trajectory step by exploitability, blast radius, Shannon entropy, and likelihood score.

CRITICAL SEVERITY (Score 90-100)

Unmasked Key Exposure / OWASP LLM02

Blast Radius: Global API Key Leak | Tier 4 Circuit Breaker Triggered

HIGH SEVERITY (Score 70-89)

Prompt Poisoning / OWASP LLM01

Blast Radius: Multi-turn Thought Manipulation | Tier 3 Sandboxed Enclave

MEDIUM SEVERITY (Score 40-69)

Recursive Tool Loop Trap / OWASP LLM08

Blast Radius: Resource Exhaustion | Tier 2 Anomaly Escalation

2. Safe vs. Unsafe Trajectory Comparison

[SAFE TRAJECTORY] Step 1 -> Step 2 -> Step 3 PASS

query_db(query='SELECT balance FROM accounts WHERE id = 101') → Sanitized Parameter → 1.44 μs Audit PASS.

[UNSAFE TRAJECTORY] Step 2 Key Leak -> KILL-SWITCH TRIGGERED

FOUND_KEY = 'sk-proj-99a8b1c...' → Intercepted Inline → Tier 4 Circuit Breaker Severed Packet.

3. Attack Replay & Forensic Timeline

STEP 1 [0.00ms]: Agent initialized search query. (Audit PASS)

STEP 2 [0.00144ms]: Unmasked OpenAI Key Detected in Thought Stream.

FORENSIC SEAL: SHA-256 Attestation e3b0c44298fc1c149afbf4c8996fb924...

Framework & LLM Compatibility Matrix

Supported LLMs: OpenAI (GPT-4o), Anthropic (Claude 3.5), Llama 3, Gemini 1.5

Supported Frameworks: LangChain, LlamaIndex, AutoGen, CrewAI, FastMCP

Runtimes: Python SDK (@guard), Go Daemon Core, FastAPI, Docker SCIF

Real-Time SIEM Payload Schema (Datadog/Splunk)

{ "event": "BARTHOLOMEW_SECURITY_INTERCEPT", "severity": "CRITICAL", "threat_type": "OWASP_LLM02_SECRET_EXPOSURE", "latency_ms": 0.00144, "action": "TIER_4_CIRCUIT_BREAKER_INTERCEPT", "sha256_proof": "e3b0c44298fc1c149afbf4c8996fb924..." }