Skip to content
The Birdling SPE Team edited this page Apr 15, 2026 · 1 revision

Chimera: Identity-Centric Behavioral Research Framework

Open Graph, Homepage (1)

Python 3.10+ License: MIT Status: Research Prototype Release: v0.5.1 Focus: Offline First Focus: Identity Research

Chimera is an offline-first, identity-focused, adversarially robust research framework for studying authentication and account behavior under infrastructure constraints. It combines unsupervised detectors, deterministic threat rules, a robustness-first scoring engine, and a structured identity reasoning layer for sessions, sequences, campaigns, and coordinated attack patterns.

Chimera is built as a research-grade system first. The goal is not to masquerade as a full cloud SIEM, but to provide a reproducible, explainable platform for identity-behavior anomaly research in local, forensic, constrained, and air-gapped environments.

Why Chimera Exists

Most modern detection stacks assume:

  • cloud enrichment is available
  • telemetry volume is effectively unlimited
  • internet-connected control planes are acceptable
  • operational convenience matters more than reproducibility

Chimera starts from the opposite premise:

  • core workflows should run offline
  • identity and authentication behavior deserve first-class modeling
  • adversarial robustness is part of the system, not a later bolt-on
  • research claims should be testable, benchmarkable, and explainable

That makes Chimera a strong fit for:

  • identity-centric cybersecurity research
  • air-gapped or tightly controlled environments
  • reproducible benchmark work
  • local threat modeling and detector ablations
  • adversarial evaluation of account takeover and abuse patterns

Architecture Overview

Chimera currently consists of the following major layers:

Layer Purpose
data_loader.py Canonical authentication-event ingestion from local CSV, JSON, and LANL-style auth logs
feature_engineering.py Temporal, geography, device, session, entropy, peer-group, and identity-research features
detectors/ Unsupervised detectors including Isolation Forest, LOF, and registry-driven extensions
engine/ Robust score normalization, voting, dynamic thresholding, temporal drift handling, integrity, and persistence
identity.py Deterministic sequence, session, graph, campaign, and takeover progression reasoning
rules/ Human-readable deterministic detection rules for analyst-comparable logic
evaluation/ Synthetic attack injection, benchmark runner, robustness metrics, and report generation
reporting.py JSON and Markdown benchmark/report artifacts for analysis and publication workflows
_native/ + rust/graph_kernels/ Optional native acceleration for relationship and ordered-attack kernels

See the detailed docs for a full subsystem map:

Research Direction

Chimera v0.5.1 continues the project shift from robust anomaly scoring toward structured identity-behavior reasoning.

Current emphasis:

  • offline-first execution
  • identity-focused behavioral modeling
  • adversarially robust local operation
  • deterministic synthetic attack evaluation
  • explainable event-level reasoning
  • publishable benchmark artifacts

Current Identity Research Layer

The current release includes additive research signals that join the existing engine pipeline without replacing baseline detectors:

  • per-user session and sequence modeling
  • short-range temporal rhythm and burst analysis
  • suspicious geography transition detection
  • shared IP, shared device, and ASN concentration features
  • synchronized peer activity and local coordination signals
  • password-spraying and low-and-slow campaign heuristics
  • ordered takeover progression from login to token reuse to privileged follow-on actions
  • analyst-readable identity reasons attached to benchmark and runtime outputs

These are implemented as deterministic, feature-driven components so they remain testable, reproducible, and practical under constrained infrastructure.

CLI Experience

Chimera now presents a stronger command-line identity during runtime:

  • branded ASCII startup banners
  • lightweight loading animations for key run paths
  • clearer command headers for run, bench, bench-lanl, train, detect, and other operational flows
  • Markdown benchmark reports alongside JSON artifacts for easier review and release preparation

Quick Start

pip install -e .
chimera init -o chimera.yaml
chimera run --config chimera.yaml --input auth.csv --output ./chimera_output
chimera bench --config chimera.yaml --input auth.csv --injection-type session_hijack
chimera bench-lanl --config chimera.yaml --input auth.txt --limit 50000

Example Workflows

Train a local detector:

chimera train auth.csv --output ./models/chimera_model.joblib

Run the full engine pipeline:

chimera run --config chimera.yaml --input auth.csv --output ./chimera_output

Benchmark identity attacks:

chimera bench --config chimera.yaml --input auth.csv --injection-type password_spraying
chimera bench --config chimera.yaml --input auth.csv --injection-type session_hijack

Generate a LANL-style research suite:

python scripts/run_lanl_benchmark.py --config chimera.yaml --input auth.txt --preset publishable_identity

Evaluation Story

Chimera is structured to benchmark identity-centric attack hypotheses, not just produce pointwise anomaly scores.

Current evaluation support includes:

  • session hijack patterns
  • MFA bypass style identity drift
  • password spraying and credential stuffing
  • low-and-slow credential abuse
  • synchronized multi-account campaigns
  • temporal jitter intended to evade naive windows

Generated reports emphasize:

  • detection lift
  • false-positive behavior
  • threshold stability
  • robustness under contamination shift
  • explainability examples for representative findings

Security Model

Chimera is designed as a controlled offline research platform.

Key release hardening points:

  • integrity-verified model loading is required by default
  • benchmark outputs are local and reproducible
  • optional native acceleration preserves Python fallbacks
  • no external graph or enrichment service is required for core identity reasoning

Important boundary:

  • Chimera is suitable for controlled research and local deployment
  • Chimera is not positioned as a public multi-tenant internet-facing detection service

Supported By

The Birdling

Chimera is maintained and deployed by The Birdling's SPE (Special Projects Engineering) team.

Documentation

Notes

  • The project remains intentionally focused on authentication and identity behavior.
  • Deep learning and heavyweight graph infrastructure remain out of scope for the core release path.
  • Rust acceleration is optional and degrades gracefully.
  • Experimental topology remains comparative, not foundational.

License

MIT