Cyber insurance covers a company’s losses from cyber incidents: the costs of response and investigation, of restoring systems and, under some programmes, business interruption and liability to third parties. Policies were built around attacks from outside: a break-in, a leak, ransomware. A mistake of the company’s own AI agent, such as deleted data or a wrong payment, is as a rule not named in these lists of risks. In 2026 the insurance market is splitting: AI risks are starting to be excluded from general policies, and separate products are appearing for them.
In brief
- A classic cyber policy covers malicious events: unauthorized access, a leak, encryption and destruction of data, long hidden presence of attackers.
- A mistake of the company’s own agent is a separate question: whether it is covered depends on the terms of the specific contract.
- United States, 2026: ISO endorsements that exclude generative AI from general liability policies; separate AI liability policies with limits of up to $25 million.
- Russia, 2026: SOGAZ and Ingosstrakh offer cyber insurance programmes; no separate policy for the mistakes of AI agents was found among the products described publicly.
- The underwriter looks at the maturity of protection, and for AI, at a tamper-protected log of the agent’s actions and human approval of sensitive actions.
What is cyber insurance?
It is insurance of the financial consequences of cyber incidents. What it covers depends on the programme and usually includes:
- response and investigation: the work of specialists who contain the attack and find the point of entry;
- restoration of systems and data;
- business interruption: lost income while the systems are down;
- liability to third parties: for example, to clients whose data leaked.
Before signing a contract, the insurer assesses how well the company is protected: with a questionnaire, an audit or an automated check of the infrastructure.
Does a cyber policy cover a mistake of an AI agent?
It depends on the contract, and this is the main question to ask the insurer before deploying agents.
The lists of risks in cyber policies describe malicious acts. For example, the SOGAZ programme “Kiberaptechka” (a “cyber first-aid kit”) names four groups of threats: unauthorized access to systems and data; exfiltration and leakage of data; encryption and destruction of data (ransomware and wiper attacks, deliberate deletion or damage of data); and cyber espionage. An accidental mistake of the company’s own agent is not named in that list.
Borderline cases:
- the agent was “talked into” something with prompt injection and exported data: whether this was unauthorized access or an action with legitimate rights is decided by the wording of the contract;
- the agent deleted data by mistake: there was no malicious act;
- the agent made a wrong payment: this is more a matter of contractual liability to the counterparty than of cyber risk.
Who answers for such mistakes, and under what rules, is described on the page who is liable for AI mistakes.
What exclusions appeared in 2026?
In the United States, since January 2026 there have been three optional ISO (Verisk) endorsements to commercial general liability (CGL) policies that exclude generative AI: CG 40 47, a full exclusion; CG 40 48, advertising and personal injury only; and CG 35 08, products and completed operations. They are not part of the base form; the insurer attaches them at its discretion. By April 2026, W.R. Berkley, Chubb, Travelers, Berkshire Hathaway and AIG had announced that they use such endorsements; by mid-2026 there were 2,369 exclusion filings in 49 states (Insurance Business, 16 September 2026).
At the same time, separate policies are appearing. In January 2026, Armilla announced an AI liability policy with a limit of up to $25 million per organization, backed by Lloyd’s underwriters; the covered categories include AI agent failures: wrong decisions, misuse of tools, escalation errors. Each such policy includes independent certification of the AI system.
What does an insurer want to see in an agent’s action log?
This is described in most detail by the AIUC-1 standard, on which insurance policies for AI agents are built (in February 2026, ElevenLabs became the first company with a policy in force based on AIUC-1 certification):
- an activity log of the AI system is mandatory (requirement E015): the full execution chains of agentic processes, including delegation to subagents and human approvals;
- protecting the log against tampering and the possibility of independent verification are mandatory; immutable storage is an additional measure (E015.4);
- human approval of sensitive tool calls is an additional measure (D003.4).
For a company, this means that a log an administrator can quietly correct proves little to an underwriter. What is needed is a log in which every action of the agent is tied to the rule under which it was admitted, and which cannot be rewritten after the fact. How a log chained with the GOST hash function works, and where its evidential force ends, is described on the page about the Streebog log (in Russian); the architecture for checking actions is on the page IS 3.0.
Are there such products in Russia?
Cyber insurance for businesses exists in Russia; as of 16 October 2026, no separate policy for the mistakes of AI agents was found among the products described publicly.
| Insurer | Product | What is known from public sources |
|---|---|---|
| SOGAZ | “Kiberaptechka”, a cyber risk insurance policy with a response and investigation service | Four groups of threats (access, leak, encryption and destruction, cyber espionage); the coverage amounts on the page are 3, 8 and 15 million roubles, or another amount; compensation of the costs of response services actually provided. Not available to financial market participants, state bodies, digital asset operators or the gambling industry |
| SOGAZ and Ingosstrakh | Cyber insurance programmes with the technology partner Yandex Cloud (announced on 4 August 2026) | Yandex Cloud assesses the security of the client’s cloud infrastructure with its Yandex Security Deck service; the client can pass the report to the insurer to calculate the policy terms. According to the press release, the volume of cyber risk protection requested by businesses under SOGAZ contracts doubled in the first half of 2026 and exceeded 12 billion roubles |
The SOGAZ requirements for “Kiberaptechka” show what an underwriter checks: no trial versions or vulnerable software, a firewall, a person responsible for cybersecurity, antivirus, a password policy for administrators, two-factor authentication for remote access, monitoring systems, staff training, protected wireless access points, compliance with the personal data legislation.
Russia’s AI law, No. 243-FZ, mentions insurance once: the developer of a sovereign or national model may insure liability for harm caused when the model is used (Article 7, point 5).
What to do
- Read your policy for exclusions related to AI and automated systems.
- Ask the insurer in writing whether a mistake of your own agent is covered, and an action of an agent misled by prompt injection.
- Keep a log of agents’ actions that is protected against changes after the fact: without it, it is hard to prove either an insured event or its absence.
- Set rules for an agent’s sensitive actions, and human approval where it cannot be done without.
- Check the contracts with model providers: who answers for a mistake, and within what limits.
Frequently asked questions
What is cyber insurance in simple terms?
It is insurance of a company’s costs and losses from cyber incidents: the work of response and investigation specialists, the restoration of systems and, under some programmes, business interruption and liability to third parties.
Does cyber insurance cover AI mistakes?
Classic cyber policies describe malicious events: a break-in, a leak, ransomware. An accidental mistake of the company’s own AI agent is as a rule not named in these lists, so the answer depends on the terms of the specific contract. Abroad, separate AI liability policies are appearing.
Is there cyber insurance in Russia?
Yes. For example, SOGAZ has “Kiberaptechka”, a cyber risk policy with a response and investigation service, and in August 2026 SOGAZ and Ingosstrakh announced cyber insurance programmes with an assessment of cloud infrastructure through Yandex Cloud.
Why do insurers exclude AI from policies?
The risk is new: there are few loss statistics, and one model error can show up at many clients at once. In the United States, since 2026 insurers have been attaching optional ISO endorsements that exclude generative AI to general liability policies.
What is needed to insure the risks of an AI agent?
To show that it is under control: a list of agents and their rights, rules for sensitive actions, human approval where it is needed, and a tamper-protected log of the agent’s actions. This is what the AIUC-1 standard requires, on which policies for AI agents are built.
Sources
- SOGAZ, “Kiberaptechka”, product page (in Russian): https://www.sogaz.ru/corporate/finrisks/cyber-pro-kit/
- Yandex, “SOGAZ and Ingosstrakh together with Yandex B2B Tech launched cyber insurance programmes for businesses”, 4 August 2026 (published by AK&M, in Russian): https://www.akm.ru/press/sogaz_i_ingosstrakh_vmeste_s_yandex_b2b_tech_zapustili_programmy_kiberstrakhovaniya_dlya_biznesa/
- Armilla, announcement of 21 January 2026: https://www.armilla.ai/resources/armilla-ai-raises-lloyds-backed-coverage-to-25m-as-traditional-insurers-retreat-from-ai-risk
- The AIUC-1 standard: https://standard.aiuc-1.com
- Federal Law No. 243-FZ of 26 July 2026, Article 7, point 5 (in Russian): https://www.consultant.ru/document/cons_doc_LAW_540336/
Checked: 16 October 2026. Insurance terms are set by the contract; request them from the insurer before buying a policy.