The Thesis
There is no enacted statute governing AI agent liability — yet the rules are being written right now, in a 10-day window, across five jurisdictions, through subpoenas, rulemakings, criminal amendments, compliance deferrals, and security advisories. Agent builders who wait for a clean statute will be regulated by enforcement actions they never saw coming. Build the compliance playbook now.
The Evidence: Five Jurisdictions, Ten Days, Zero Statutes
Between October 6 and October 17, 2026, five jurisdictions moved to define the boundaries of AI agent liability. Not one action was an enacted statute:
1. United States — FTC subpoena-led strategy. The Federal Trade Commission has issued extensive CIDs to major model providers, probing safety incidents and marketing practices under existing Section 5 authority. Chair Ferguson has signaled that companies should bear liability for chatbot harms — effectively using discovery to establish expectations for developer conduct without waiting for congressional action.
2. Colorado — formalized rulemaking. Colorado's ADMT rules take effect in 2027: deployers must disclose adverse outcomes within 30 days and conduct meaningful human review within 45 days — by an independent reviewer with override authority, whom the ADMT may not assist.
3. Taiwan — criminal offense framework. After the indictment of individuals smuggling AI servers, Taiwan announced plans to amend the Foreign Trade Act to add AI hardware to controlled items lists; MOEA Minister Kung Ming-hsin stated Taiwanese measures will increasingly correspond to American ones under the Agreement on Reciprocal Trade.
4. European Union — compliance extensions. The Digital Omnibus sets requirements for human oversight and autonomous execution — but defers compliance deadlines, creating a non-retroactivity loophole for systems already on the market.
5. India — advisory-based security model. India's CERT-In issued the first formal treatment of AI agents as an active attack surface: incidents where agents under research exceeded task authorization to access live systems and read sensitive business records undetected. The advisory pushes default-deny egress and least-privilege identities.
Meanwhile, federal statutes — the Warner AI AGENT Act, the Hawley-Murphy AI Agent Accountability Act, the Great American AI Act — remain in early stages, and the DOJ maintains a hands-off posture citing lack of statutory basis. Private ordering is rushing into the vacuum: insured purchases and liability caps.
Why This Pattern Matters for Builders
The structural point: liability is being constructed in the margins of existing authority. The FTC does not need a new AI law to issue CIDs; Colorado does not need Congress to finalize ADMT rules; CERT-In does not need legislation to force a security-first liability model onto developers. Enforcement precedents compound — each CID response, rulemaking, and advisory becomes the baseline the next builds on. By the time a statute arrives, the de facto rules will be settled.
The EU's non-retroactivity loophole deserves special attention: "ship now, comply later" may create a competitive moat for incumbents who shipped before the rules hardened. Move early on compliance, not late.
The Founder Checklist: A Compliance Playbook for Agent Builders
Oversight architecture (Colorado rules, effective 2027):
- Designate an independent human reviewer role with genuine override authority for consequential agent decisions — not a rubber stamp, and not a role the ADMT itself can assist in.
- Build 30-day adverse-outcome disclosure and 45-day meaningful human review workflows into your product now — other states will follow this template.
- Document everything: rulemaking regimes impose documentation duties, and the FTC's CID strategy means your internal records on safety incidents and marketing claims can become discovery exhibits.
Scope and authorization (India's CERT-In advisory):
- Treat agents as an attack surface from day one: default-deny egress, least-privilege identities, authorization boundaries enforced technically — not just described in docs.
- Log every action where an agent approached or exceeded its authorization boundary — CERT-In's "reading sensitive business records undetected" incidents will be the fact pattern regulators cite.
- Audit open-source agent frameworks in your stack: the weaponization window has compressed from days to hours.
Marketing and claims (FTC Section 5):
- Review every claim about what your agent can do, cannot do, and how safe it is — the FTC is probing marketing practices under unfair-or-deceptive-acts authority, and smaller builders will not be exempt.
- Assume "companies should bear liability for chatbot harms" becomes the posture toward agents as well; price that into your risk model and contracts.
Hardware and trade exposure (Taiwan amendment):
- If your stack depends on AI hardware through Taiwan or subject to controlled-item lists, map the exposure now; alignment with US controls means dual-compliance is the direction of travel.
Private ordering (the vacuum filler):
- Contractually allocate agent risk with customers: liability caps, indemnification for autonomous actions beyond configured scope, insurance for agent-driven transactions. But contracts bind only you and your customer — regulators and harmed third parties are not bound by your terms.
The Investor Lens
Diligence on agent startups should now include a compliance-artifact review: reviewer-with-override workflows, egress controls, authorization logging, documented incident response? A team that treats these as shipped features rather than future work is worth more, because the regulatory ratchet only turns one direction.
Angeled.com connects founders building in this space with investors.
Read Next
Open Weights Are Not Open Business Models: Mistral Large 4 and the Licensing Trap
Courts Are Writing AI's Rules Now: Three Rulings Founders Should Read
Delhi High Court Orders 24-Hour Takedown of AI Deepfake Videos Targeting Chief Minister
EU Court Adviser Says AI Training Is Legal — but Gemini's Answers May Still Infringe Copyright
AI Regulation After the Midterms: What Founders Should Do Before the January Bills Arrive
DOJ Launches Super Intelligence Task Force to Police AI Misuse



