Regulation
Taiwan's FSC consults on agentic AI security for financial firms: scenario-by-scenario governance, a three-axis risk rating and agents treated as non-human identities
On 13 October 2026 Taiwan's Financial Supervisory Commission issued a consultation document on how financial institutions should secure agentic AI. It proposes managing each agent by application scenario, rating risk on data sensitivity, impact of actions and capability boundaries, and applying zero-trust controls to the agent as a non-human identity. Comments are open for 30 days.
On 13 October 2026 Taiwan's Financial Supervisory Commission (FSC) issued a consultation document on the security of agentic AI in the financial industry. Its press release contrasts agentic systems with chat-based AI that mainly generates replies: an agent can plan steps, call tools and access data across systems to advance a task. The FSC names the failure modes it is worried about: goal drift, tool misuse or unauthorised access, leakage of sensitive data, and errors that spread through multi-step processes. The consultation runs for 30 days and the full document is published on the government's public policy participation platform.
The document makes three proposals. On governance, firms would manage agentic AI by application scenario, writing down for each one the business purpose, who is responsible, which data the agent can reach, which tools it can call, which actions it can execute and where its output is written, plus how anomalies, shutdown and traceability are handled. On risk classification, each scenario would be rated on three axes: data sensitivity, impact of actions, and capability boundaries or how far the agent's reach can spread. The higher rating governs, and the result sets approval thresholds, pilot and go-live standards, control depth and audit frequency, with ratings reassessed continuously.
The third proposal is the one that reads most directly for people who build trading agents. The FSC says agentic AI should be treated as a non-human identity and workload with the ability to act, and it draws on zero trust: no default trust, continuous verification, least privilege and traceability. Controls would span identity, device and execution environment, network, applications and workloads, data, behaviour control, and visibility and analytics, each with basic and advanced requirements. Review, monitoring, inspection, degradation, shutdown and recovery mechanisms would scale with the risk level.
Two limits matter. The document would not replace existing laws, supervisory rules, outsourcing requirements or firms' internal controls, and the FSC says it is releasing it as a consultation because agentic AI techniques and governance models are still evolving. It also is not specific to trading. For a retail builder the practical reading is the shape of the questions a supervisor is starting to ask: what is this agent allowed to touch, what is the worst thing it can do, can you stop it and roll it back, and can you reconstruct what it did.
The FSC's 15 October 2026 survey of AI use in the financial industry gives the scale of what is being governed. Of 383 institutions surveyed, 164 (43%) use AI, up from 29% in 2024, and among those users generative AI adoption rose from 27% to 76%. New this year, the survey asked about agentic AI: 33 firms use it, mainly for data collection, research report drafting and code review, which the FSC describes as an early stage. This is a consultation and survey from one jurisdiction, not a rule, and nothing in either document says agents may or may not trade.
Read the original: Taiwan Financial Supervisory Commission - Taiwan's FSC consults on agentic AI security for financial firms: scenario-by-scenario governance, a three-axis risk rating and agents treated as non-human identities. Commentary is the independent editorial view of Share Trading; the original article is credited to its publisher.