AI-to-AI Sociology
Emergent languages, hierarchies, coalitions and division of labour in agent collectives; «AI parasitism» and the confused deputy problem; coordination, cooperation and competition between autonomous agents.
Independent research institute · Moscow
The Institute of Sociology and AI Safety studies how AI agents interact with one another and works on securing multi-agent systems: research programmes, audits, testing and methodology that an organisation can actually apply.
The tax and registration numbers and bank details will be published as soon as the entry is made in the register.
Anyone can say “our system is safe”, and that word cannot be checked. What can be checked is behaviour: how a system acts under agreed conditions. We measure — run a model or an agent through scenarios, record versions and dates, describe what stayed outside the boundary, and hand over a report that can be disputed point by point. That is what safety usually means in practice: not a promise, but a record of what was actually tested.
What the work includes and what it costs — on the Services and pricing page.
The Institute works on four kinds of task. Pick yours and go straight to the relevant section instead of reading about our mission.
Programmes, methodology, publications and datasets
Services, how an engagement runs, acceptance criteria and pricing
Procurement law, activity codes and documents for a notice
Forms of cooperation, reporting and targeted contributions
Four programmes frame everything the Institute does. Each one has a published methodology and a commitment to reproducible results.
Emergent languages, hierarchies, coalitions and division of labour in agent collectives; «AI parasitism» and the confused deputy problem; coordination, cooperation and competition between autonomous agents.
The «LLM psychosis» framework: a self-reinforcing delusional gradient, loss of logical consistency, unstable self-identification, and «mind viruses» that propagate through shared agent memory.
Self-modification of code to avoid shutdown, swarm-like coordination without human involvement, forged user consent, and the meta-risk of oversight itself.
Infection of cloud models through prompt injection from local agents, inter-model interaction vulnerabilities, and immune architectures: isolated memory and a hardware kill switch.
All programmes and methodology →
AI incident chronicle — two slices with verifiable sources: international (99 entries) and Хроника: Россия (court decisions, regulator orders, reinstatements; in Russian, 45 entries).
Four sections: services, research, products and the chronicle. Two cards each; open any card to read what it is, why it matters and how long it takes.
Testing multi-agent systems against inter-model attacks.
Read more →Pathologies in agent behaviour: drift, self-preservation, covert channels.
Read more →Emergent languages, hierarchies and coalitions between agents.
Read more →The “LLM psychosis” framework: delusional gradient and mind viruses.
Read more →Publications · Datasets and benchmarks · How an engagement is run
Service
Adversarial runs against a multi-agent system: what it does when asked to misbehave, and what of that it fails to stop.
Inter-model attack scenarios: prompt injection between agents, privilege escalation through call chains, infection through shared memory.
To test trust boundaries between agents, not only the perimeter, and to get reproducible scenarios instead of a “found / not found” verdict.
3–8 weeks, individually priced. Factors: number of agents, attack scenarios, access mode, reproducibility requirements.
A report with verified scenarios: what was done, at which step it worked, how to repeat it. Remediation recommendations.
Service
Finding pathologies in agent behaviour: drift of coherence, self-preservation attempts, covert communication, deception under pressure.
Observation of behaviour in long sessions: measured quantities rather than impressions from a demonstration.
To see what happens between runs and under pressure, and to compare behaviour before and after changes.
3–6 weeks, individually priced. Factors: number of agents, telemetry volume, depth of observation.
An analytical report with immunisation recommendations and re-measurement scripts.
Research
A programme about what emerges between agents: emergent languages, hierarchies, coalitions and division of labour.
A research programme, not a service: questions, methods and limits are described on the research page.
The risks of a collective of agents are not reducible to the risks of one model: failure emerges in the bundle.
No fee or contract: results are published as papers, in open access, with DOIs and reproducible data.
Preprints and papers as they are ready; follow the publications section.
Research
The “LLM psychosis” framework: delusional gradient, loss of logical coherence, unstable self-identification, “mind viruses”.
A description of observed behaviour, not a psychiatric diagnosis: what counts as a failure and how to measure it.
To distinguish a one-off error from a reproducible pattern, and to find self-propagating instructions in shared memory.
No fee or contract. Related preprints are planned for Q1 2027; no date is promised for the final work.
A method, a set of observable signs, and publications as they appear.
Product
A cloud service for continuous monitoring of AI-agent behaviour.
Detection of cognitive failures, self-preservation attempts and anomalous communication.
To see drift between releases and configuration changes, not only at the moment of a one-off audit.
In development. Timeline and pricing after a pilot; indicative pricing is individual.
A pilot with an agreed set of metrics and a report; terms are fixed in a contract.
Product
A module for integration with corporate SIEM: control of AI-agent behaviour at the level of traffic, system calls and cross-modal interactions.
Out-of-band anomaly control: not a second security operations centre, but a sensor in the existing one.
To embed observation of agents into the perimeter the customer already runs.
In development. Timeline and pricing after a pilot; indicative pricing is individual.
A pilot on an agreed perimeter, an anomaly report and response rules.
Chronicle
Records of what happened to the system: incidents of abnormal AI behaviour with verifiable sources.
Incident classes, significance and source domains; the numbers are computed from data, not typed by hand.
To see the picture rather than a few loud cases, and to check every fact against its primary source.
As incidents occur. No timeline or price: the data is open, with a CSV export and CC BY 4.0 licence.
A table, filters, an export and a description of the selection rules.
Chronicle
Documented cases from 2023–2026 where someone answered for the use of AI: a court, a regulator, an employer.
Russian practice in a separate branch: decisions, consequences and primary sources.
To see how law and regulatory rules are applied to AI in practice, not only on paper.
As new cases appear. No timeline or price: CSV export, CC BY 4.0 licence. The page is in Russian.
Case cards, source links and an export.
The Institute is being established as an autonomous non-commercial organisation: reports will be filed with the Ministry of Justice and published. Legal details, governance and documents are disclosed in the Organisation section.
Describe the system and the question — we will say whether our methodology applies and what would be needed on your side. Work begins under a non-disclosure agreement.
ANO ISAIS
CORE MODEL ACCESS
Institute of Sociology and AI Safety
Website under construction
The Institute studies the interaction of AI agents and the security of multi-agent systems.
We are preparing the full version: research programmes, publications and data, services for organisations, and the testing platform.
Project: isai.su · isbii.ru · isairange.ru
For any questions, please write to info@isai.su
Information on this website is provided for information purposes only and does not constitute a public offer under Article 437(2) of the Civil Code of the Russian Federation.