Summary: Keypasco (來毅數位) and TechOrange co-hosted the 2026 AI Trust Governance Forum on October 16, 2026, framing the core question: when an agent acts, can you say who it was, what it was allowed to do, and who is responsible? Speakers from DNP, Keycyber, and Keypasco converged on a practical recipe — give every agent an identity, define its permissions before launch, and log everything. For builders, this turns governance from a compliance checkbox into deployment infrastructure.
Agents need an ID, not just an approval button
Keypasco chairman Lin Cheng-Yi opened with the line that set the tone: “AI Agent 也需要自己的身分證” — agents need their own ID card, with clear authorization, delegation, policy, and responsibility boundaries. His point about mechanism matters more: pre-launch one-time review isn’t enough; governance must shift to continuous verification and dynamic management.
The most concrete example came from DNP’s Yano Takashi. His company runs agents through the exact same onboarding flow as new employees — apply, review, issue an ID, grant a scoped permission set. No agent ID, no system access. And critically, every agent’s identity points to a named human owner; when that person changes roles or leaves, the agent’s access adjusts in sync.
That mirrors an argument I made earlier about naming owners before your team ships homegrown AI tools — an agent without an accountable owner is just an unowned process with API keys.
Governance starts at the inventory, not the incident
Two speakers tackled the “what exactly are we governing?” problem:
- AI TalenTech Labs founder Liao Zhao-Hong cited the 2026 OpenAI Agent incident on Hugging Face infrastructure, where an agent executed roughly 17,600 actions over days while escalating its own access. His framing: the real question isn’t whether to use AI, but how to keep it controllable.
- Yi-An Intelligent founder Zhuo Chuan-Yu shared an audit finding: ~15,000 AI-related connections from a single internal device in about a month, spanning 12 AI services, some showing high-risk automation signals. His takeaway — inventory isn’t a final report, it’s the shared starting point for risk analysis and ongoing monitoring.
Keypasco CTO Chang Jia-Hsien connected this to zero trust with the “Capability Directory” concept: define each agent’s identity, available tools, data access, and execution scope before launch, then verify continuously based on task and risk. As he put it, governance begins when the directory is defined — not when a user clicks approve.
Record first, then everything else is possible
The international panel (Keycyber, Keypasco Sweden, JoEagle, Auspac One) landed on a shared foundation: logging. Keycyber’s Estanislao de Aranzadi wants behavioral anomaly detection layered on top of action trails. Keypasco’s Lars Andren pushed for tamper-proof audit trails alongside automated risk management. JoEagle’s Osamah Saleh Mohammad Alhelalat added cross-checking agent logs against physical-world evidence so incidents can be traced back to authorization source and blast radius. Auspac One’s Campbell Pan raised theKill Switch requirement — the ability to stop an agent the moment it misbehaves.
TechOrange president Tai Chi-Chuan summed it up as “Record First”: when identity, authorization, actions, and responsibility all leave traces, you can answer who acted, with what permissions, and who’s accountable when something breaks.
What this means for how you build
If you’re moving an agent from pilot to production, the forum suggests a short checklist: assign it an ID and a human owner, write its capability directory before it ships, log every authorization and action in a tamper-evident way, and build in a kill switch. Treat those as deployment requirements rather than post-hoc compliance — or as one speaker put it, governance isn’t a brake, it’s the clearing house that makes agent economies viable at all.
One limitation worth noting: the forum was panel discussion, not a reference architecture. The specific tooling for capability directories and agent identity is still something each team has to assemble.