glacis-io/autoredteam

How glacis-io/autoredteam works

autoredteam is an LLM security evaluation product: users configure a target and launch an attack campaign, then inspect scored, evidence-backed results and optionally harden defenses or emit governance policy. The principal runtime stages are probe generation, target execution, scoring, evidence capture, and reporting; the campaign model is the shared internal data layer. The README documents broader hardening, validation, provider, and training capabilities. Some runtime wiring is unsampled, so only documented or excerpt-supported relationships are shown.

Main entry points: campaign_runner.py Campaign runner · autoharden.py Defense hardening · base.py Provider session · campaign.py Campaign model

Components 25

Campaign workflow

CLI cli.py
Campaign runner campaign_runner.py
Campaign model campaign.py
Trajectory execution trajectory_engine.py

Targets and probes

Attack generation attack.py
Pack registry attack_packs/registry.py
Pack interface attack_packs/base.py
Domain packs attack_packs/domains
Target setup prepare.py
Provider registry providers/registry.py
Provider session providers/base.py
Cloudflare provider providers/cloudflare_workers.py

Results and governance

Scoring and judging scoring.py
Judge model models/judge.py
Evidence attestation attestation.py
Report generation reporting/generator.py

Hardening and training

Defense hardening autoharden.py
Defend loop defend.py
Defense learning immune.py
Policy emission emit_policy.py
Training data training_data.py
LoRA training training/lora_trainer.py

Other components

Engineer
LLM target
Glacis service

Connections 23

Engineer → CLI: launches
CLI → Campaign runner: runs campaign
Campaign runner → Campaign model: uses model campaign.py
Pack registry → Pack interface: registers packs attack_packs/registry.py
Pack registry → Domain packs: loads built-ins attack_packs/registry.py
Pack registry → Campaign model: builds campaign attack_packs/registry.py
Attack generation → Campaign model: models probes
Target setup → Provider registry: selects provider
Provider registry → Provider session: registers sessions providers/registry.py
Cloudflare provider → Provider session: implements session providers/cloudflare_workers.py
Provider session → LLM target: sends requests
Trajectory execution → Provider session: sends turns trajectory_engine.py
Trajectory execution → Campaign model: records traces trajectory_engine.py
Campaign runner → Scoring and judging: scores results
Scoring and judging → Judge model: uses local judge scoring.py
Campaign runner → Evidence attestation: records evidence
Campaign runner → Report generation: writes report
Defense hardening → Defend loop: uses defenses autoharden.py
Defense hardening → Defense learning: uses learning autoharden.py
Defense hardening → Policy emission: emits policy autoharden.py
Defense hardening → Training data: collects examples autoharden.py
Training data → LoRA training: trains adapters training/lora_trainer.py
Evidence attestation → Glacis service: submits hashes

Drawn by GitDiagram on November 10, 2026 from the repository's file tree, README and source files. Read as Markdown ·