At what point does an AI task cross the line: Tiandun Zhi'an implements task-level security contracts

When enterprises use AI to organize customer data, analyze contracts and call business systems, a task that originally only required "reading and analysis" may gradually involve data export, external transmission or record modification. A single compliant response does not mean that continuous execution is always within the authorized scope. The more tools AI can call, the more enterprises need to answer a specific question: To what extent is the AI allowed to proceed in this task?

TianDun ZhiAn is developing an enterprise AI security middle platform, planning to build an independent security control layer between enterprise AI applications, data and business systems. Focusing on three types of risks: sensitive information leakage, non-compliant AI behavior and cross-border multi-round tool calls, it implements enterprise rules into the authorization, execution and audit process of tasks.

Clarify task boundaries first, then constrain each step of operation

The product idea of TianDun ZhiAn comes from the team's attention to the actual enterprise usage environment. Different enterprises have different data classification, system permissions and approval processes, so security requirements need to penetrate into specific businesses, rather than just staying in a set of general content filtering rules.

According to the R&D plan, the system first combines user intent, task context and data sensitivity level to form a "task-level security contract", clarifying available tools, accessible data and maximum action limits, and then verifies permissions, parameters and approval conditions before tool invocation. For identified risks, disposal measures such as desensitization, rewriting, approval or blocking will be taken in accordance with enterprise policies.

For example, in a customer data analysis scenario, "reading specified records and generating internal summaries" and "exporting all customer data and sending it to external addresses" belong to different authorization scopes. TianDun hopes to make the system not only understand the semantic difference between the two, but also implement restrictions on specific tools and parameters. This is a scenario to explain the product mechanism, not a completed customer case.

On this basis, the platform design further connects model calls, tool parameters, external requests and execution results, and continuously checks whether multi-step actions conform to the original goal. Its focus is to identify operations that seem reasonable in a single step but cross the boundary when combined, helping enterprises extend security management to the entire execution process of the Agent, and retaining traceable basis for alarms and manual intervention.

Enable security policies to update with business, and make updates verifiable

Enterprise systems and business processes will change, and security policies also need continuous maintenance. For this purpose, TianDun ZhiAn has designed a policy iteration process with three Agents working in coordination: Policy Builder converts materials such as enterprise systems and business specifications into candidate rules; Eval Builder constructs normal, non-compliant and boundary samples; Optimizer is responsible for evaluation, rewriting and retesting.

The candidate policy will take effect after being confirmed through testing and release, and retain version traceability and degradation rollback mechanisms. False positives, false negatives and new business demands found in subsequent product use are planned to be fed back to the rules and evaluation samples to support the next round of updates.

This solution aims to reduce the burden on enterprises to write and maintain complex rules, while controlling the impact of rule adjustments on normal business. Task-level authorization, cross-step judgment and controlled policy update together constitute the product direction of TianDun ZhiAn for business Agents. The specific interception effect, false positive level and response performance still need to be tested in platform construction and actual scenario verification.

According to project materials, the team has submitted a domestic invention patent application for "A dynamic management and audit system for enterprise AI usage security based on multi-layer semantic recognition".

Cut in from enterprise scenarios to explore continuous services

In terms of market expansion, the team plans to focus on industries such as software services, intelligent manufacturing, healthcare and pharmaceuticals, finance and insurance, connect with IT, security and business leaders who have clear AI application scenarios and security needs, and make gradual progress from demand communication, scenario assessment to product verification.

The commercialization plan includes revenue from POC and deployment projects, fixed subscription revenue and over-limit usage billing, and takes into account the needs of enterprises for private deployment and targeted adaptation. The team hopes to precipitate reusable capabilities in delivery and provide services around the ever-changing AI usage needs of enterprises. Current enterprise contacts are still in the early stage of demand communication, and no feedback based on actual product use has been formed yet.

Complementary experience in enterprise-level engineering, AI R&D and commercialization

The team members' experience corresponds to technology products and enterprise architecture, product and commercialization, semantic model and evaluation respectively, providing the project with R&D and demand understanding foundations from different dimensions.

TianDun ZhiAn is currently in the stage of platform R&D and construction. The funds are self-raised by the founding team, and it plans to launch a seed round financing of 4.5 million RMB, which is mainly used for team building, product development, market promotion and basic operation. Next, the team will focus on supplementing R&D and delivery forces, promote core function construction and enterprise scenario verification, and gradually form an AI security platform suitable for the landing environment of Chinese enterprises, so that security rules can penetrate into every data usage and business action of AI.