Filed under ai-agents
Every brief tagged "ai-agents".
- § 01 · AI-AGENTS
Fault for Services, Strict Liability Only for Machines: Wang Liming on Tort Liability When AI Agents Cause Harm
In Tsinghua University Law Journal (《清华法学》) 2026 No. 5, Wang Liming (王利明) of Renmin University sets out how Chinese tort law should handle harm done by AI agents. His answers: an agent is never a liable party, so liability falls on the people along its value chain; AI that provides a service, such as generative AI, is judged by fault, while strict product liability applies only to agents that enter the physical world as tangible things and threaten life, body, health or tangible property; where fault applies, operators and deployers should prove they took reasonable care; multiple actors are not jointly and severally liable by default, but answer in proportion to their causal contribution, with the party closest to the end user carrying the heaviest risk-control duty; open-source and free providers get lighter treatment; causation should be presumed when the evidence sits with the defendant; and the 'mass micro-torts' agents produce, such as leaks of many people's personal information, call for representative and public-interest litigation rather than individual suits. DCC summarizes the article section by section and maps each proposal against the Supreme People's Court's AI Disputes Opinions (法发〔2026〕10号), the PIPL and the Civil Code. Most of Wang's framework is already in the Opinions; the reversal of the burden of proof for service agents, the presumption of causation, and substitute measures for data-subject rights go further.
- § 02 · AI-AGENTS
Who Lets a Personal AI Agent In? Meta Muse, US Banks and Retailers, and the Pre-Research Behind China's Agent Data-Processing Standard
A full translation. 网安寻路人 opens a series, written with the core drafting group of the national standard Security Requirements for Data Processing by AI Agents (智能体数据处理安全要求), now being drafted, which shares the standard's preliminary research. The first installment is a case study, not a Chinese rule. It follows how US banks and retailers responded after Meta launched its personal agent Muse on September 8, 2026 and passed 10 million users in about six weeks. Amazon, JPMorgan, Schwab, and Morgan Stanley blocked or restricted outside agents, while Walmart, Shopify, and Best Buy opened controlled gateways. The author traces two fault lines. The first is identity, overreach, and liability: agents logging in with customers' credentials leave platforms unable to tell the customer, a mistaken agent, and an impostor apart. The second is control of shopping entry points and profit pools, from Amazon's roughly $70 billion ad business to banks' low-cost deposits. The industry's answer is moving from 'silent agents' toward verifiable agent identity (KYA), scoped credentials, read-only interfaces, and human approval for high-risk acts. With the original's seven charts, its timeline, its ten-row comparison of institutional policies, and its 26 notes.
- § 03 · AI-AGENTS
Keeping an AI Agent Inside Its Personal-Information Boundaries: What Meta's PrivacyLib Shows About Making PIPL Purpose Limits Work at Runtime
On October 27, 2026, Meta described PrivacyLib, a layer of its privacy infrastructure that enforces privacy rules where backend services actually read and write data. AgentLib ties each agent tool call to its session, so the check happens even when an agent strings tools together in ways no developer anticipated. 网安寻路人 uses the disclosure to answer a question the PIPL leaves open for agents: how do you hold a system whose execution path unfolds at runtime to Article 6's purpose and minimum-necessity limits and Article 13's legal bases? The author's answer has five parts. Turn the user's task into concrete processing boundaries, set by trusted business systems rather than the model. Check every operation where data actually flows, not at the final answer. Carry provenance and use conditions onto derived content (summaries, vectors, long-term memory), and re-classify when inference yields sensitive personal information. In agent-to-agent delegation, grant only the intersection of the upstream's delegable scope, the sub-agent's own permissions, and the subtask's needs; keep control instructions apart from business data; and map each tool to its legal role under Article 21 (entrusted processing) or Article 23 (provision to another handler). And make the rules themselves current, revocable, and auditable. Five diagrams from the original are included.
- § 04 · FIVE-YEAR-PLAN
What the 15th Five-Year Informatization Plan Signals for AI and Data Compliance Through 2030
数据何规 (What Rules for Data) reads the Central Cyberspace Affairs Commission's 15th Five-Year Plan for National Informatization, published October 23, 2026, for what it means for AI and data compliance over the next five years. Foundational AI legislation and judicial interpretations on AI and personal information are now formal tasks, though the law's name, form, and timing are not settled. On AI regulation, the plan pairs an unexplained 'optimization' of generative AI filing with a new security risk assessment system for applications in key sectors, and with assessment and filing for AI agents, on-device large models, and anthropomorphic services. The author reads that as a shift from model-level to use-level scrutiny. On data, the author reads cross-border facilitation as clearer boundaries and lower costs for the three existing export routes, not their abolition, and notes that it runs alongside the CAC's special inspections of data export. The plan also singles out personal information testing of smart devices and biometric information management. The author's bottom line: a plan creates no new obligations, but it marks where the next rules will land.
- § 05 · COMPUTE-CENTERS
One Machine Room, Five Regulatory Identities: MaaS Compliance for China's AI Compute Centers
AnJie Broad partners Cai Hang and Yao Ting and associate Liu Zeqiang argue that the IDC-era compliance checklist no longer fits the AI compute center. Their thesis: one machine room now carries five regulatory identities at once — domestic IDC operator, cross-border AI service provider, service exporter under the Export Control Law, supplier of self-deployed model capability, and responsible entity for the agents it ships — each with its own logic, and each transmitting obligations to the others. The brief works through value-added telecom licensing (B11/B12), energy-conservation review and PUE caps, the continuing duties under the Regulation on Network Data Security Management (State Council Decree No. 790), the token-export exemptions in CAC Decree No. 16 and what they do not exempt, the territorial limits of the GenAI Interim Measures, the Export Control Law catch-all, the unsettled line between large-model filing (备案) and large-model registration (登记), open-weight license trigger clauses, and the criminal exposure that follows an agent's tool calls under Criminal Law Article 285.
- § 06 · AI-AGENTS
Tokens Meter Input, Not Value: Hong Yanqing on Beijing's Agent Measures (Part 4 of 4)
Part 4, closing Hong Yanqing's commentary on the Several Measures of Beijing Municipality on Accelerating Agent-Led Development (北京市关于加快智能体引领发展的若干措施, 京发改〔2026〕1185号). The Measures' Article 6 proposes a Token (词元) economy — Token-as-a-Service, Agent-as-a-Service, Results-as-a-Service, and a shift from billing by Token consumption to value-based billing; Article 8 funds 'Token factories' and Token vouchers. Hong draws the line the policy still needs: Tokens measure the consumption of intelligent means of production, not the value of intelligent products. Tokenization differs across models; a task's full cost includes tool calls, memory storage, human review, and failed retries; and Token volume has no fixed ratio to task value — so treating Token throughput as industrial performance rewards long contexts, loops, and retries. His alternative is a five-layer evidence chain (resource input → system capability → valid task results → process results → enterprise and social value), a cost-per-valid-completed-task formula that counts review, retries, and expected risk losses, and an attribution discipline of pre-launch baselines and phased pilots. Outcome billing must be corrected for quality and risk — narrow metrics make customer-service agents rush calls and procurement agents chase price cuts, and vendors can cream-skim easy tasks while humans absorb the hard residue — so projects with unstable task boundaries should blend base, resource, and performance fees rather than jump to pure Results-as-a-Service. Different policy objects need different-layer metrics, mapped onto Part 3's five maturity levels, and fiscal support should pass staged evidence gates: prototypes may fail, pilots must beat baselines in real business, demonstrations must replicate at acceptable cost, and commercial-stage projects must survive subsidy taper — with prompt exit for projects that stop producing new evidence.