Liability for AI: so far it is written only by those it will bind
The issue for 23 October. There are no special rules of liability for AI — not in the EU, not in the US, not in Russia: general law applies, and that is what courts use. The special rules are written by those they will bind. Four stories in two weeks — and one conclusion.
1. US banks cannot tell an AI agent from a customer
22.10.2026. Major US banks are recording activity of unauthorised third-party AI agents inside customer accounts. Agents sign in with the customer's login and password, in some cases call call centres, and "take actions not covered by the account holder's direct instructions". Bank executives call the situation "scary", Axios reports as retold by Izvestia.
What matters here is not the scale but the cause: banks still cannot distinguish an agent's actions from a person's. So the response is technical rather than legal — block such technologies and rewrite service rules. Morgan Stanley has not recorded unauthorised access to its internal systems, but has already begun blocking access to its platforms for such agents.
Why this is about liability. Identity systems rest on the assumption that only a human knows the login and password. An agent formally satisfies that assumption: it arrives with the correct credentials, therefore it is the customer. When the agent errs, nobody is liable under general rules: the customer granted access, the bank broke no procedure, the agent is not a legal person. There is no special rule that would allocate this risk.
And this is not a hypothesis about the future. The gap is already in money: while a bank cannot tell an agent from a human, the legal question "who acted" cannot be answered, and only a technical answer remains — ban it.
Source: Axios, 22.10.2026, as retold by Izvestia — iz.ru. We did not open the original Axios report — that is a caveat, not a formality.
2. Anthropic writes a law about itself — and edits it after criticism
22.10.2026. Anthropic is rewriting part of its own draft federal AI safety bill after a wave of criticism from AI safety specialists and legislators. The trigger was a confidential document the company discussed with congressional staff and which Politico published. Critics say the proposed rules are weaker than the bipartisan Frontier Act and would make it harder to block dangerous models.
Under the leaked text, authorities could seek to prohibit a model only under a narrow set of conditions: the risk must be predictable and immediate, and the harm must exceed 1 billion dollars or involve the death of, or serious injury to, more than 50 people. Grounds would be bioweapons development, a model escaping its developer's control, or a cyberattack. A ban would run only through a federal court at the Commerce Department's request; the draft forbids other agencies from suspending a model on their own.
Separate objections concern limits on state authority and requirements for independent auditors: evaluators could be paid by the developers themselves, with no common licensing standards. On 18 October Anthropic's head of federal policy Anthony Cimino acknowledged that wording on federal preemption of state laws was "not precise enough" and promised to revisit risk thresholds and the approach to blocking. California State Senator Scott Wiener said the proposal does not protect public safety sufficiently; criticism also came from the other side, from former White House AI adviser David Sacks.
Worth noticing. The company is not lobbying someone else's bill — it brings its own. And it draws criticism from both directions at once: from safety advocates "too soft, a dangerous model cannot be stopped" and from industry "too harsh, it will choke development". Even with honest intent, a threshold of 50 deaths is a threshold written by a company, not by the people who die.
Source: IXBT citing Politico, 22.10.2026 — ixbt.com. The risk thresholds must be checked against a second independent source before publication — they are the key figures of the issue.
3. European Parliament: the scrapped liability directive may come back
21–22.10.2026. Centrist parties supporting Ursula von der Leyen clashed at a plenary session over holding AI companies liable. The European Commission considered special liability rules in 2022 but abandoned them in 2025 in favour of simplifying procedures and boosting competitiveness.
Vice-chair of the Socialists and Democrats group Alex Agius Saliba called the withdrawal "a big mistake" and urged the Commission to table the bill. Left-centre and Green MEPs demand a new law and liability for model developers by analogy with manufacturers of ordinary goods. Right-centre MEPs oppose it: lower administrative burden and more private investment. Commissioner for International Partnerships Jozef Sikela recalled that the rules were withdrawn because EU countries and MEPs could not agree.
A short timeline without which the dispute makes no sense: in 2022 the Commission proposes the AI liability directive, in 2025 it withdraws it in the name of simplification, in October 2026 Parliament demands it back. Two years to prepare, one year to scrap, now back again. That is the price of an unresolved question: while there is no rule, the risk sits with whoever is caught first — the bank, the doctor, the customer.
Source: Politico, as retold by Vzglyad, 22.10.2026 — vz.ru.
4. Russia: security requirements instead of liability
We covered FSTEC order No. 117 of 11.04.2025 in the fifth issue: it has been in force since 1 March 2026, it has two working clauses about artificial intelligence — 60 and 61 — and we separately showed that the "direct ban on cloud AI services" was attributed to the order by news outlets rather than written in its text. We will not repeat that here.
For this issue's subject, one thing follows from the order: it answers "how to protect information and the system", not "who compensates harm when the model errs". Clause 61 requires that a human take the decision based on the model's output — but that is a duty of the operator, not a rule under which the model developer answers. Liability is simply absent from the order: not "weakly drafted", but outside its subject.
And that is a telling picture. Russia builds regulation from system security: 243-FZ sets duties — labelling, model admission, access to data — and order No. 117 sets information protection requirements. Neither answers who pays when a neural network causes harm. In Europe the question was at least recognised as a question — and the directive was withdrawn. Here it has not been raised yet.
The order covers the public sector — state information systems, systems of state bodies, state unitary enterprises and state institutions; it does not regulate the commercial sector. The title and dates were verified against the primary source — publication.pravo.gov.ru, publication No. 0001202506170011.
5. What follows
- The absence of a directive is also a decision. The EU did not forget about liability — it scrapped it. While there is no rule, the cost of an error falls on whoever has the least choice: the bank customer, the patient, the reader.
- Self-regulation produces text, not independence. Anthropic's draft is useful, yet the 50-deaths threshold was written by a company. The question is not honesty but the author's position.
- Technical indistinguishability of agent and human is a new gap, and it is already in money. While banks cannot tell one from the other, the question "who acted" cannot be answered.
- A practical takeaway for us: in services and at the test bench, record that the decision on test results is taken by a person and keep the trail of who pressed the button. That is not a guarantee of liability, but it is the only thing that works while the rules are absent.
6. The issue's figures and where to check them
| Figure | What it means | Where to check |
|---|---|---|
| 22.10.2026 | publication date of the US banks AI agents report | Izvestia citing Axios |
| $1 bn | harm threshold in the Anthropic draft | IXBT citing Politico |
| 50 | threshold of deaths or serious injuries in the same draft | same |
| 2022 · 2025 · 2026 | EU liability directive proposed, scrapped, demanded back | Politico via Vzglyad |
| 8 | states suing OpenAI under consumer protection law | fifth issue |
| 60 and 61 | clauses of FSTEC order No. 117 that deal with AI | fifth issue, order text |
| 01.03.2026 | the order has been in force since this date | publication.pravo.gov.ru |
| 11.04.2025 | date of FSTEC order No. 117 | same |
The issue was prepared on 23.10.2026. External sources were opened on the day of preparation; retellings are marked in the text. If you find an error, write to info@isai.su — we will fix it and mark the correction.