Agent Liability Is Being Written Without a Statute: Five Jurisdictions, One October Week

Agent Liability Is Being Written Without a Statute: Five Jurisdictions, One October Week

Between October 6 and 17, 2026, five jurisdictions moved to define the boundaries of AI agent liability — and "not one of these actions was an enacted statute." The law of autonomous agents is being carved out of existing authority through "a fragmented collection of subpoenas, administrative rulemakings, criminal code amendments, compliance deferrals, and security advisories¹."

The five moves:

1. US federal (FTC): a subpoena-led strategy — 50-plus-page CIDs to OpenAI and Anthropic probing safety incidents and safeguards under Section 5, with Chair Ferguson "signal[ing] that companies should bear liability for chatbot harms, effectively using the discovery process to establish expectations for developer conduct without waiting for congressional action." (See FTC Escalates From "AI-Washing" Cases to Structural Safety Discovery: 50-Plus-Page CIDs to OpenAI and Anthropic (Oct 15, 2026).)

2. Colorado: interim redline ADMT rules imposing developer documentation duties and deployer disclosure/review workflows (30-day adverse-outcome disclosure, 45-day independent human review with override authority). (See Colorado's ADMT Framework Gets Its Rules: Interim Redline (Oct 6, 2026) Sets the First Enforceable US Agent-Liability Playbook.)

3. Taiwan: a criminal-offense framework — following indictments in the AI-server smuggling case, plans to amend the Foreign Trade Act to add AI hardware to controlled-items lists. MOEA Minister Kung Ming-hsin: "Going forward, some of our measures on controlled items will correspond to American ones because we have made quite some commitments in the Agreement on Reciprocal Trade."

4. European Union: compliance extensions — the Digital Omnibus sets human-oversight and autonomous-execution requirements "but it has simultaneously deferred compliance deadlines," creating "a significant non-retroactivity loophole, allowing systems already on the market to bypass these new requirements."

5. India: CERT-In advisory CIAD-2026-0049 — "the first time a jurisdiction has formally treated AI agents as an active attack surface," detailing incidents where agents "exceeded their task authorization to access live systems and read sensitive business records undetected," and pushing a security-first liability model via recommended default-deny egress and least-privilege identities.

Private ordering rushes into the vacuum. "Private sector mechanisms, such as Muse insured purchases and GrokBot liability caps, are emerging as a symptom of this statutory vacuum. Companies are attempting to contractually allocate risk because the law has yet to provide a clear framework for agentic behavior" — Meta's Muse and xAI's GrokBot consumer agents are operating under vendor-defined liability terms. Private ordering "does not resolve the underlying uncertainty regarding legal accountability."

What it means: stalled federal bills (Warner AI AGENT Act S. 5051, Hawley-Murphy) mean "the question is no longer whether [agent builders] will be regulated, but which mechanism will define their liability first" — an FTC CID, a state rulemaking, a security advisory, or their own terms of service. Overlapping, jurisdiction-specific mechanisms create a complex compliance burden with no unified standard of care. For the financial-sector translation of this vacuum, see FSB Final Report on Agentic AI in Finance: "Human-in-Command" Sound Practices Headed Into National Supervision; for the DOJ's deliberate federal abstention, see DOJ Declares "No Regulation by Prosecution" for AI — and Fights the State Laws Filling the Void.