Founded and led by HKAAST · Secretariat provided by HKAAST Greater Bay Area · 3 Locks: Safe · Reliable · Trustworthy 繁 · 简 · EN Want to know how the National AI Safety Governance Framework 3.0 lands in your industry?
GBA AI Safety & Development Standing CommitteeNon-profit AI Safety Think Tank

Open-Source Tools

Part one lists tools incubated by the Committee's seven research programmes; part two is a separate column recommending third-party community tools for everyday self-protection (not developed by the Committee).

1. Our tools — incubated from research programmes

Under the principles of openness, auditability and self-hosting. Tools marked "research preview" release their source code on an "open after academic publication" schedule; licence types will be announced at release. Institutional partners may join development via the Secretariat. Each tool lists its licence plan, self-hosting requirements and use cases, so you can assess the fit for your environment.

Behaviour steering · Agent Manager

HoloManagerResearch preview

The agent butler (new flagship): actively steers and influences LLM/agent behaviour — goal anchoring, challenge escalation, hypergraph fleet governance. Embeds the HoloGovernor monitoring layer, forming a "steering + monitoring" double-layer governance. The steering wheel steers; the speed limiter + dash cam watch.

  • Licence plan: behaviour-steering layer in development; release schedule announced at launch
  • Self-hosting: standard Python environment, layered on top of the HoloGovernor monitoring engine
  • Use cases: multi-agent fleet governance, goal-drift correction, challenge escalation review
  • Framework 3.0 mapping: agent governance page
Side-channel monitoring · Governor

HoloGovernorCore engine open-sourced

The agent monitoring engine: millisecond risk scoring and three-segment triage for every operation (release / block / deep audit), a tamper-evident audit chain supporting independent third-party replay, and periodic governance reports — the embedded monitoring layer of HoloManager.

  • Licence plan: core engine open-sourced (GitHub, alexlihk/hologovernor)
  • Self-hosting: standard Python environment, local pip install; community edition provides a minimal closed loop
  • Use cases: agent operation guarding, auditable evidence chain, audit cost tied to risk distribution
  • Code: GitHub alexlihk/hologovernor (community edition: alexlihk/aegis)
  • Online demo: 3token.org/demo/
Guardrail

HoloAegisResearch preview

Zero-shot LLM content-safety manifold: frozen representations + topological inference deliver a minimal-parameter, auditable content guardrail with alert mode; 40 automated tests passed.

  • Licence plan: source code released on the open-after-publication schedule; licence announced at release
  • Self-hosting: standard Python environment; CPU suffices (frozen representations, minimal parameters)
  • Use cases: content-safety gates for open LLMs, community/support moderation, alert-mode risk warnings
  • Live demo: 47.76.62.228/demo/aegis/ (demo environment, account via the Secretariat; demo firewall pack 3.2MB, deployed P50 26.3ms)
  • Whitepaper: not yet public — about two weeks
From research: arXiv 2608.08485
Circuit breaker

HoloBlockerResearch preview

Embodied-AI topological circuit breaker: computes the topological distance between behaviour and the safety manifold in real time; past the threshold it cuts power in milliseconds — for industrial internet, low-altitude economy and agent applications.

  • Licence plan: source released on the open-after-publication schedule; core components open-sourced first
  • Self-hosting: integrates with the execution layer / hardware, deployed in partner co-build mode
  • Use cases: industrial robots, drone swarms, financial and consumer-facing agents
  • Whitepaper/technical docs: GitHub shennong-ai/HoloBlocker
From research: ChinaXiv 202604.00098
Geometric audit

HoloAuditorResearch preview

Maps domain axioms (e.g. the TCM "monarch–minister–assistant–envoy" logic) onto mathematical manifolds and audits AI output geometrically in real time — intercepting "popular but wrong" answers in 0.82 ms, raising data quality and cross-border regulatory mutual recognition.

  • Licence plan: source released on the open-after-publication schedule; licence announced at release
  • Self-hosting: standard Python environment; domain-axiom mapping modules tailored per scenario
  • Use cases: data-quality audit and regulatory mutual recognition in TCM and other expert domains
  • Whitepaper/technical docs: GitHub shennong-ai/HoloAuditor-Demo
From research: SSRN 6368538
Privacy audit engine

SA-RI Audit EngineReports on request

Split-learning / split-inference representation-leakage measurement: dual attack families, cross-metric empirical audit, AMBER/RED-graded export-safety reports; measurements report variance and boundary conditions — unmeasurable cases are marked INCONCLUSIVE.

  • Licence plan: core engine v0.3.x is open source (BSD, GitHub); audits commissionable now
  • Self-hosting: Python + PyTorch; a single consumer GPU completes the full audit protocol
  • Use cases: pre-export assessment of split-learning deployments, vendor due diligence
  • Whitepaper/engine docs: GitHub alexlihk/splitrisk (v0.3.x)

📄 Sample export-audit report (Track C · GitHub) →

From research: RG 10.13140/RG.2.2.32481.67681
Data space

Trusted Data Space SuitePartner co-build

For industrial data spaces such as food & TCM: onboarding/listing dual review, controlled-task registration, metering and reporting — aligned with the "Data Element ×" strategy and the cross-border "Data Customs" concept.

  • Licence plan: per cooperation agreement, partner co-build licensing
  • Self-hosting: containerised deployment in partner environments (incl. SM-crypto modules), with manuals
  • Use cases: food & TCM data spaces, cross-border data-compliance pilots
  • Whitepaper: not yet public — about two weeks
Partner cooperation project · via the Secretariat
Public education

AI Safety AssistantOpen to all

The Committee's public-education chatbot built on our knowledge base: phishing, passwords, privacy, deepfakes, data export — free, instant, bilingual.

  • Licence plan: online service free to use; knowledge-base content may be quoted for education
  • Self-hosting: pure Python standard library; runs on any OpenAI-compatible LLM endpoint
  • Use cases: schools, community groups, public libraries and other education settings
  • Whitepaper: not yet public — about two weeks
Try it now →
─ Third-party community tools below (not developed by the Committee) ─

Recommended open-source tools for everyday self-protection, maintained by global communities and independent of the Committee's research programmes; listed for public-education purposes only.

Password managersKeePassXC / Bitwarden — "one strong password per account", the most cost-effective defence (open source, self-hostable).
Browser protectionFirefox + uBlock Origin — blocks trackers and malicious ads, shrinking the phishing surface.
Breach checkHave I Been Pwned — check whether your email appears in known breach dumps; change passwords promptly.
Two-factor authAegis Authenticator / hardware keys — always enable 2FA on important accounts.
📌 Open source is our methodology, not a slogan: a tool's rules, thresholds and test results should all be independently verifiable. Trust comes from verifiability, not black boxes.