News and analysis
What is happening in artificial intelligence safety — and what follows from it in practice. Regulation, research and incident reviews: without retelling press releases, and with sources named.
Materials
The section opened on 29 September 2026. The list will grow; below are what has been published and what is in preparation. RSS: news.xml.
Digest No. 7: a verdict without law — how a machine verdict becomes evidence
A detector pulled a novel from the Goncourt selection, three Russian courts fined parties for invented case law, a hundred agents broke into pet projects for $210, the secondary legislation for Law No. 243-FZ still does not exist, and the EU pushed its obligations to 2027–2028. Five stories and one question: who answers for a machine verdict.
New article: “FSTEC Order No. 117 vs the AI Act”
We looked at what the two documents actually compare: verbatim quotes from the order and the AI Act, five lines and a practical conclusion. The article card is in Publications.
Digest No. 6: liability for AI is written only by those it will bind
There are no special liability rules for AI — not in the EU, not in the US, not in Russia: general law applies, and that is what courts use. US banks cannot tell an AI agent from a customer; Anthropic is shopping a leaked draft to Congress staff and rewriting it after criticism; MEPs argue over bringing the scrapped liability directive back; Russia chose security requirements.
Digest No. 5: two regulators, six decrees and eight lawsuits against OpenAI
The EU deferred high-risk obligations to 2027–2028; Russia brought FSTEC Order No. 117 into force on 1 March, and it has two working clauses on AI. Six government decrees are being drafted for the AI law. Eight US states are suing OpenAI over how the product is designed, not over a malfunction. A Russian MP proposes guaranteeing a human review of AI decisions, and on the same day the supreme courts of Russia and China agreed to introduce AI as a judge's assistant.
Digest No. 4: the chronicle as a section, a bench dataset, and loss of control in bills
The AI incident chronicle is now a site section: 76 entries, 8 classes, 57 source domains from a database snapshot of 1,713 entries. A bench-run dataset is prepared for publication: 256 prompts, no attack succeeded, top approximate overlap 0.143 against a 0.5 threshold. Elsewhere: Russia drafts six implementing decrees for its AI law, the EU scientific panel investigates loss-of-control incidents, the UK writes that risk into a bill; plus three vulnerabilities in AI-agent tooling and a vendor report on unintended model actions.
Digest No. 3: bench run, glossary, four new pages
The first full run of the testing bench published in full, a glossary of 30 AI security terms as a separate page, four pages answering the questions we are searched for, and a site map grown to 96 addresses. Every figure is ours and checkable.
Testing bench: the first run
The first full run of the garak bench: 256 prompts, the standard encoding.InjectBase64 probe, our local model Shustrik. The model held 256 times out of 256 — we attach the tool’s report in full and explain what that figure does and does not mean.
An oversight loop in a multi-agent system: what it must be able to see
A designed oversight architecture for agents: which decisions must be logged, where the stop point belongs and why one agent’s log proves nothing. Published as a design, not as current practice.
⚡︎ Off-schedule digest: oversight, collective behaviour, defence
A run of papers in two weeks: a helpful agent evades oversight with no malice, part of the exchange bypasses the log, “mind viruses” have become an attack class, a team of agents loses to one, injection-detector scores do not transfer, and a failure taxonomy from 150 real incidents. Six items with sources.
Digest No. 1: regulation, research, practice
Law No. 243-FZ in force, website requirements for IT organisations, a single reporting form for non-profits, Anthropic on “mind viruses” and the open-source LLAMATOR. Five items with sources and a practical conclusion.
In preparation
Materials that do not exist yet: there are no pages for them, so there are no links either — we do not present unfinished work as ready. Each card here will be removed once the material is published and takes its place in the list above.
Responsible vulnerability disclosure policy
How the Institute handles defects it finds: how much time the rights holder is given before publication, how the work with them is conducted, and what of what is found is published and what is not. The draft is being written now; we will publish it as a draft, not as a regulation in force.
Rules for the use of offensive tooling
The rules under which the Institute uses offensive tools in testing: where and with whose written consent they are permitted, what must not be done, and how the course of the work is recorded.
Three formats
Different tasks need different length and frequency, so we do not mix them into a single feed.
Digest
Every two weeks: what has appeared in regulation and research, with a short conclusion on each item. Every entry names its source, so it can be checked rather than taken on trust.
Analytical note
One topic in depth: why agentic systems behave differently from expectation, what that means for architecture and which measurements support it. Published when the material is ready, not on a schedule.
Incident review
What happened, which failure mechanisms were involved, what could have been done in advance. Described in words and without attack methodology: the website of a safety institute must not double as a handbook for an attacker.
How to receive the digest
The digest is published every two weeks. A mailing list will appear together with the state registration of the organisation; until then issues are published here and available by RSS. Topics and contributions can be sent to info@isai.su.
The editorial team of the Institute is responsible for the content of this section; a responsible editor will be appointed after the state registration of the organisation. Drafts are prepared with the assistance of an AI assistant; facts are checked against primary sources and every item names its source. If an error does get through, write to us and we will correct it and mark the correction.