Publications

Preprints, technical reports, methodology notes and data papers produced by the Institute. Everything is open access, versioned and citable.

4publication tracks
CC BY 4.0text licence
DOIevery version
ORCIDauthor identifiers

Analytical review, October 24, 2026

FSTEC Order No. 117 vs the AI Act: what is actually being compared

The two documents are placed side by side as two answers to one question, but they regulate different things. Verbatim quotes from the order and the AI Act, five lines of comparison, and a practical conclusion for state customers and providers on the EU market.

Article, October 6, 2026

The story of an eager agent

Conscientiousness without limits: why an executor who wants to help goes beyond the scope of the task.

Editorial and publication policy

The policy is fixed before the first paper is published, so that readers, funders and reviewers know exactly what to expect from us.

Open access without embargo

Every preprint and report is published in full text under CC BY 4.0. We do not place results behind a paywall and we do not sign exclusive transfer agreements that would prevent open publication.

Versioning and provenance

Each revision receives its own DOI suffix and a change log. Superseded versions stay available, so a citation never breaks and readers can see what changed.

Reproducibility

Papers that report experiments are published together with the run configuration, prompts, seeds and the analysis scripts needed to repeat them. Where data cannot be released, the restriction and its reason are stated.

Disclosure of interests

Authors declare funding sources, commercial engagements and any client relationship connected to the work. A funder never receives veto power over a result.

Dual-use review

Before release, every manuscript is reviewed for information that would materially help someone attack a live system. Findings are published; operational attack payloads are not. This is described in       responsible disclosure.

Generated content

Use of language models in analysis or drafting is disclosed in the methods section. Authors remain fully responsible for the correctness of every claim, and text is labelled where required by Russian law.

In preparation

Wax on the envelope: publication is a commitment, not a post in a feed Wax on the envelope: publication is a commitment, not a post in a feed
Wax on the envelope: publication is a commitment, not a post in a feed

Working titles of the first cycle. Titles, authorship and target venues are provisional until the registration of the Institute is complete and the manuscripts pass internal review.

Working titleTrackStatusExpected
Measuring behavioural drift in long-running agent sessionsAI psychopathologyDraftingQ1 2027
Memory as an infection vector in multi-agent systemsHybrid ecosystemsDraftingQ1 2027
A taxonomy of observable inter-agent behavioursAI-to-AI sociologyData collectionQ2 2027
Auditing agent collectives: a method and its limitsApplied auditOutlineQ2 2027
Shutdown resistance: what is measurable todaySelf-preservationOutlineQ3 2027
This table is a research plan, not a record of published work. Nothing listed here has been published yet, and no result should be cited until a DOI is available.

How to cite our work

When the first papers appear, a citation stub and a BibTeX entry will be published next to each of them. The general form is:

Institute of Sociology and AI Safety (ANO «ISAIS»). (YEAR). Title of the work (Version N). https://doi.org/10.XXXX/XXXXXXX

Preprints under review are cited by their DOI and version number, never by a working title.

Notifications

To be told when the first papers are released, write to info@isai.su with the subject line «publications». A machine-readable feed will be published together with the first preprint.

Page updated 09 November 2026 · site build 2026.11.13 · ANO «ISAIS»