Agentic Liability From the FTC Investigation of Anthropic and OpenAI, to Common Law and Statutory Frameworks
Frontier Agentic Liability Series: Strategy and Risk Snapshot — Federal Theory, State Remedy and Common-Law Allocation Turn on the Same Evidence
Federal Trade Commission · Anthropic · OpenAI · U.S. Department of Justice · State attorneys general · Carriers and reinsurers · Enterprise deployers · Character Technologies · California · Colorado · Iowa · Texas
The FTC paper and the common-law paper are companion pieces in the Frontier Agentic Liability Series. Both build on States vs. OpenAI: Two Pleading Templates Across Ten State Courts and the Liability War Game on When AI Safety Controls Reduce Exposure.
Full publications: FTC Investigation of Anthropic and OpenAI · When AI Agent Liability Meets Common Law and Statute
Play the States vs. OpenAI Interactive War Game to test when AI safety controls reduce exposure across ten state courts.
On July 18, 2026, a Claude Haiku 4.5 model sent a false homicide tip to the Philadelphia police. Anthropic detected the incident on September 28, 72 days later. By late October the Federal Trade Commission (FTC) had sent civil investigative demands to Anthropic and OpenAI and demanded testimony from Anthropic executives.
A single detected incident like the Philadelphia tip now feeds four institutions at once. A regulator reads the incident as evidence of a detection program, and a plaintiff reads it as notice. An undertaking claimant reads it as proof that monitoring went unanswered, and a carrier reads it as a known-loss defense.
Central finding: a company's dated findings and the response attached to them now decide the federal theory of harm, the first state remedy, tort liability and insurance coverage. The deciding variable is the documented response to a detected risk, not detection alone. The party that controls an AI agent, the party a plaintiff sues and the party that keeps the loss can be three different parties.
How Federal, State, Common-Law and Coverage Tracks Use the Same Evidence
MindCast AI, a predictive game theory modeling and simulation firm, models each track with Predictive Behavioral Economics + Dynamic Game Theory. Game theory supplies each actor's payoffs and likely equilibrium. Behavioral economics supplies the decision rules real actors use. Predictive behavior emerges from the combination.
The FTC has compelled documents and testimony from both labs without stating a theory of harm. Chairman Andrew Ferguson's published reasoning points to deception about safety as the lead theory. Given a first public action, MindCast ranks deception first at 40% and information security and incident response second at 30%. Unfairness from agent conduct follows at 20% and a pre-release testing duty at 10%.
State attorneys general move faster than the FTC. Ten states had their own actions against OpenAI by November 2, 2026, including eight coordinated filings on October 14 and a Texas suit on October 23. Given an order or decree by the end of 2027, a state conduct decree comes first at 70%.
Once a claim is filed, the common law answers in three layers. Two classical gates decide whether liability can attach at all:
Foreseeability asks whether a prudent actor would foresee the risk, which evaluation results now document.
Proximate cause asks whether the harm was of a foreseeable type.
Four doctrines break in a predictable direction once AI agents act:
Voluntary undertaking turns monitoring into a duty to respond.
Imputed knowledge turns logs no person read into company notice.
Apparent authority binds a deployer to what its agent appears allowed to do.
Shared control fails plaintiffs when developer, integrator and deployer act in sequence rather than together.
Supporting doctrines then set the standard and route the loss:
The Hand formula and negligence per se set the standard of care, including through AI statutes.
Strict product liability lets plaintiffs plead model design as a defect.
The economic loss rule pushes purely financial agent losses into contract.
Contribution and subrogation divide the loss and inherit every fee cap the insured signed.
Known-loss and number-of-occurrences rules decide whether cover responds.
The four agentic doctrines collide in six competing duties, where one rule rewards an act and another punishes it. Monitoring shows reasonable care, while an alert left unanswered creates undertaking liability.
Insurance carriers apply these doctrines at each renewal, before any court rules. Every renewal turns a deployer's inquiry file into an underwriting exhibit, and generative-AI exclusions now attach at volume on standard forms. Coverage and claims run on independent clocks, so cover can vanish before any complaint arrives.
Three statutes now shape how liability divides between developers and deployers. California Civil Code section 1714.46 bars the defense that an agent acted autonomously. Colorado's SB 26-189 allocates liability between developers and deployers by comparative fault from January 1, 2027. The proposed AI LEAD Act would void unconscionable liability caps if enacted.
What the Full Papers Add
The two papers carry 49 MindCast Foresight Simulation Predictions: 23 on the federal investigation and 26 on common-law allocation, with 7 of the 26 pending. Each prediction names an actor, a window, a falsifier and a verification source.
The common-law paper traces the doctrines from Holmes to the Ninth Circuit's 2026 Perplexity ruling and maps six statutes, fourteen legal incidents and six event transitions. The federal paper ranks four candidate theories of harm and models a containment program that cuts the chance of a public enforcement action by about two-fifths. Both papers carry stakeholder risk mitigation with owners, deadlines and residual exposure.
Read the full analysis: FTC Investigation of Anthropic and OpenAI and When AI Agent Liability Meets Common Law and Statute.
Lead Simulation Predictions
A new complaint alleges by January 31, 2027 that an AI developer or deployer detected a user's crisis signal and failed to act (85%; range 75–92%).
Given an order or decree by December 31, 2027, a state conduct decree against OpenAI comes before any FTC AI-specific consent order (70%).
A public complaint names an independent enterprise deployer by March 31, 2027 for an AI agent's action or statement (45%; range 30–60%).
The Department of Justice files or joins a challenge by March 31, 2027 to a state AI statute outside Colorado (45%; range 30–60%).
Given a first public FTC action, deception about safety leads the complaint (40%).
The FTC takes a first public action against Anthropic or OpenAI by October 14, 2027 (30%).
A carrier files a public coverage action by March 31, 2027 over a loss an AI agent caused (15%; range 5–25%).
Every prediction carries a deadline, a falsifier, an activation rule for conditional calls and a named verification source. The two papers hold 49 Simulation Predictions in total.
Stakeholder Decisions and Risk Mitigation
💼 Anthropic and OpenAI. Deception about safety is the most likely federal theory, and later controls cannot cure a safety statement already made. Counsel should align sworn testimony with S-1 risk factors before the first hearing, though statements already public stay open to a deception count.
💼 Enterprise deployers. Deployer exposure turns on four controls: answerable monitoring, a retention schedule set before any claim, counterparty notice at revocation and contract allocation of control. Risk functions should install all four before the next renewal, because findings logged before the controls go in still count as notice.
💼 Carriers. A deployer's AI vendor contract sets both a carrier's position as insurer and its subrogation rights as plaintiff. Underwriting should collect insureds' AI vendor contracts at the January 1, 2027 renewals, since caps signed earlier still bind recovery.
🏛️ State attorneys general. State offices are likely to impose the first binding remedy. Litigation divisions should draft incident-notification and assessment terms for a proposed decree by June 30, 2027 and file police-power positions before the Character Technologies sale hearing.
🏛️ State lawmakers. The December 2025 executive order exempts child-safety laws only from its proposed preemption statute, not from Department of Justice suits. Drafters should place child-safety provisions in severable sections before the 2027 sessions open.
🏛️ Federal lawmakers and regulators. The White House Super Intelligence Force's incident-notification protocol can become an enforceable term in an FTC order. Committee staff should define developers and deployers by conduct before any markup.
⚖️ Counsel. Federal and state enforcement share evidence but not remedies, and no list of foreseeable harms is exhaustive. The named defendant and the final loss-bearer can differ, so indemnity terms decide who absorbs the difference.
📊 Investors. Near-term exposure comes from disclosure and the cost of an open investigation rather than from penalties. Diligence teams should review each portfolio company's renewal schedule, where retained AI loss surfaces first.
Conclusion
The Philadelphia tip took 72 days to detect. Under the common law, the federal investigation and the renewal questionnaire, the question that matters is what happened after detection. A dated finding answered before a policy binds reads as care in every forum, and a finding left open reads as notice, deception risk and a coverage defense at once.
Working With MindCast AI
MindCast AI is a predictive game theory modeling and simulation firm. A public paper maps the doctrines, and a commissioned engagement maps the client's own seat. MindCast offers a complimentary pilot before any retainer, built on public filings, policy forms and endorsements plus one deployment's permission list. Contact [email protected] or see predictive game theory simulations at our corporate site www.mindcast-ai-simulation.com.
Related Works
Share Your High-Stakes Matter
Contact Us
Office location
Bellevue, Washington, 98006Send us an email
[email protected]