A proposed framework for governed autonomous cyber defense: action envelopes, model-excluded admissibility, bounded effects, reversibility and evidence an external reader can recompute.
The proposed category and formal framework
A proposed framework for governed autonomous cyber defense: action envelopes, model-excluded admissibility, bounded effects, reversibility and evidence an external reader can recompute.
10.5281/zenodo.23089594The LLM and its defensive architecture
A proprietary domain-specific language model within a governed defensive architecture. Execution authority stays outside learned output; actions are bounded, independently observed and recorded in recomputable receipts.
10.5281/zenodo.23089652The model narrates outcome-verified sessions. It does not authorize execution. Scheduled review and element-level reconciliation close the loop after action.
Containment, verification and sealing in the reported episode
Turns in an externally recomputed receipt
Evaluation families: policy, envelope and evidence
Deployment observations are historical instances from the manuscript, not latency guarantees, statistical benchmarks or independent validation of every construction.
Original published editions. The files are mirrored without modification; checksums are provided in the public manifest.
Listings are discovery paths, not endorsements or peer review.
Explore how the papers connect defensive autonomy, language models, adversarial inputs and verifiable infrastructure effects.
How the DAI domain-specific LLM relates to cybersecurity, information security and incident response: model-excluded authority, bounded containment and verifiable evidence.
Read the research guide →Autonomous Defensive Artificial Intelligence (ADAI): a research framework for governed cyber defense, adversarial inputs, bounded incident response and verifiable effects.
Read the research guide →No. The manuscript describes a domain-specific LLM and a defensive architecture, with model output excluded from execution authority.
The papers specify the mechanisms and recomputation contract. Some properties require inspecting construction or provider capabilities; a receipt does not prove every property. DAI §8.10 explicitly discusses target-provenance limitations.
These editions are research preprints. Formal results depend on the stated assumptions; reported observations are specific to the experiments described. No third-party scientific endorsement is asserted.
The public repository hosts research manuscripts, citations, a technical glossary and the research website. Public access does not grant a license to proprietary source code or model weights.