Who Pays When AI Agents Err: Insurance for AI Risks in Russia and the West
A typical 2026 scenario involves a company connecting an AI agent to its CRM system to automatically clean duplicate contacts. The agent errs in classification, treats half the customer database as duplicates, and deletes it. No hacker attack occurs and no malicious intent exists on the developer side — the harm stems purely from autonomous execution of an erroneous model output.
Who bears responsibility? The model developer, the integrator who assembled the agent, or the company that deployed it? More importantly, does any party hold an insurance policy that would cover the loss?
Russian Legal Framework
Russian legislation consistently assigns responsibility to a human or legal entity. The National Strategy for AI Development until 2030 and the AI Ethics Code establish that liability must rest with a person or organization. The Code recommends allocating responsibility across lifecycle participants according to their roles, yet this remains guidance rather than binding law.
No dedicated civil-law norm governs AI liability. Harm caused by AI is currently evaluated under general provisions of the Civil Code. Paragraph 18 of Supreme Court Plenum Resolution No. 1 defines a source of increased danger as any activity that creates elevated risk of harm due to the impossibility of full human control. Legal doctrine debates whether an AI agent qualifies; if it does, liability arises under Article 1079 regardless of fault.
Insurance concepts appeared as early as 2020. The Concept for Regulating AI and Robotics until 2024 explicitly listed insurance of liability and compensation funds as mechanisms for compensating AI-related harm. It also noted that foreign markets often require insurance policies as a condition for market entry and proposed a “red button” for system shutdown plus a “black box” for recording incident circumstances.
Mandatory insurance already exists inside experimental legal regimes. Since January 2025, participants in such regimes must insure liability for harm to life, health, and property, including harm caused by AI-driven decisions. This covers unmanned trucks, drones, delivery rovers, and telemedicine. Conditions are set by each regime’s program.
Federal Law No. 243-FZ of 26 July 2026 “On Support for the Development of Artificial Intelligence Technologies in the Russian Federation” grants developers of sovereign or national foundational models the right to insure property interests and liability arising from model use. The provisions defining sovereign and national models enter into force only on 1 March 2027. At the same time, Article 8 obliges developers to maintain technical documentation containing key parameters and model limitations — precisely the information insurers need for risk assessment.
Draft Law vs Final Text
The March 2026 draft “On the Fundamentals of State Regulation of AI Technologies” contained a detailed responsibility chain in Article 11. All participants would be liable proportionally to their degree of fault. Developers, operators, and service owners would be released if they took “exhaustive measures” to prevent unlawful outcomes. Users would remain liable only for intentional acts or violations of usage terms. Operators who compensated victims could seek recourse from developers for hidden model defects.
The draft also addressed situations where no one is at fault: if harm resulted from hidden system defects or unforeseeable third-party actions, both operator and user would be released. In such cases insurance and compensation funds would become essential, yet the draft omitted any mention of insurance mechanisms.
The final law reduced Article 11 to a single sentence stating that participants “bear responsibility in accordance with the legislation of the Russian Federation.”
International Experience and Market Products
Neither the United States nor the European Union imposes mandatory AI liability insurance. Article 26 of the EU AI Act requires users of high-risk systems to maintain oversight, monitoring, and logs for at least six months, but does not mandate insurance. The draft AI Liability Directive was withdrawn in summer 2025.
In the US, liability arises from common law, state statutes, and case law. Texas enacted TRAIGA on 1 January 2026, while Colorado replaced its stricter law with a softer version effective 2027. California’s AB 316 took effect the same day, preventing defendants from arguing that “the AI itself” caused the harm.
Specialized insurance products already exist. Armilla (Lloyd’s) offers up to $25 million aggregate coverage for hallucinations, model degradation, AI-agent actions, and regulatory fines. Munich Re aiSure provides parametric coverage up to €15 million triggered when predefined accuracy thresholds are breached. Testudo (Lloyd’s) covers generative-AI liability up to $9.25 million. HSB (Munich Re) embeds coverage for bodily injury, property damage, and privacy or copyright claims into SME policies.
The AIUC-1 standard from Artificial Intelligence Underwriting Company defines approximately 50 requirements and 130 controls specifically for AI agents. Certification includes third-party audits and extensive agent testing; agents from ElevenLabs have undergone more than 5,000 such tests.
Conversely, major carriers including AIG, W.R. Berkley, Great American, Chubb, Travelers, and Berkshire Hathaway have begun inserting AI exclusions into general liability policies. OpenAI reportedly secured only $300 million in coverage against potential claims reaching billions.
Related articles
I Know but Cannot Tell: The Forensic Expert's Struggle to Share Real Cases Under NDA Constraints
A cybersecurity journalist reflects on the inherent difficulties of writing about digital and financial forensics while bound by strict confidentiality agreements. The article explores how verifying financial documents involves far more than spotting Photoshop alterations, raising questions about data origin, accuracy, interpretation, and the validity of conclusions drawn by examiners. Automated verification systems often produce conflicting results because they assess different aspects of a document, yet lack transparent justification for their verdicts. Attempts to anonymize cases, obtain client consent, or use synthetic data each carry limitations that can still lead to identification or fail to demonstrate actual expertise. The author recounts collaborating with an AI to draft the piece, only to realize the meta-discussion about the inability to publish real cases itself became the only publishable story. Ultimately, the piece highlights the tension between the professional desire to share knowledge and the contractual and ethical barriers that prevent meaningful disclosure in the field of forensic analysis.
Three Hypotheses on Information Security Risk Assessment: Where Automation Truly Saves Time at Avito
Avito expert Irina Sozinova details how the company tested automation in its risk assessment process after scaling issues emerged with thousands of risks in the GRC system. The team formulated and piloted three hypotheses focused on automating typical risks, eliminating manual team preparation, and validating data to reduce InfoSec workload. Results showed a 16% drop in total risks registered, zero preparation time for teams, and improved coverage and context quality, though overall expert time savings were partially offset by new post-processing tasks. The pilot confirmed that automation works best on repeatable scenarios with stable data sources rather than promising immediate proportional time cuts. Avito plans to complete automation of context collection and finding classification this year to expand risk assessments across the company next year.
Integrating OAuth 2.0 with PDP Architectures Without Modifying Core Authorization Engines
The article explores how to combine OAuth 2.0 delegated access with traditional PEP/PDP authorization systems. It proposes an OAuth-aware PDP layer that translates access tokens into standard PDP checks while preserving the original PDP implementation. Three main use cases are analyzed: stored user grants, tenant-level administrative grants, and scopes carried directly inside tokens. The design relies on synthetic subjects such as u+c and t+c pairs, ephemeral policies for token-embedded scopes, and optional OAuth-aware PAP components. Special attention is given to token contents, policy storage via PAP, and the need for deterministic subject identifiers. The approach supports both classic scope models and structured RAR-style authorizations while maintaining separation between OAuth grants and subject permissions.
Russian Businesses Face 50-Ruble Fee per MVD Query for Passport and Residence Checks
The Russian government has backed a proposal requiring businesses to pay 50 rubles for each completed session querying Ministry of Internal Affairs databases. The fee covers checks on Russian passport validity, residence registration, residence permits, and foreign citizen migration records. Projected annual budget revenue reaches 120 billion rubles. Companies must still perform client identification under Federal Law 115-FZ and the Law on Communications, with fines up to 0.1 percent of own funds for banks or 500,000 rubles for telecom operators. Smart Engines highlights its Sherlock anti-fraud platform as a local alternative that detects over 600 forgery indicators, including deepfakes and NFC chip tampering, without repeated paid queries. The company cites a June 2026 legal opinion from the National Council of Financial Markets confirming that such tools can satisfy compliance requirements.