Services and software development
Audit, testing and hardening of AI systems. All work is delivered under a contract with a fixed scope, timeline and set of deliverables.
Service catalogue
Pricing is scoped individually: there is no single price list, because the price depends on the task. The factors behind the calculation are published openly and listed for every service. We send the indicative range and the stage breakdown on request — within one business day, without registration on this site.
| Service | Scope | Deliverable | Timeline | What drives the price |
|---|---|---|---|---|
| AI Red Teaming as a Service | Testing a multi-agent system: cross-agent prompt injection, privilege escalation through call chains, infection through shared memory files | Report with reproducible attack scenarios and a remediation plan | 3–8 weeks | Number of agents and test scenarios, environment availability, access mode, requirements for a reproducible report |
| Behavioural Audit of AI Ecosystems | Behavioural assessment: delusional gradient, self-preservation attempts, covert communication | Analytical report and immunisation recommendations | 3–6 weeks | Number of agents, telemetry volume, depth and duration of observation |
| AI Supply Chain Security Assessment | Supply chain review: open-source components, datasets, LoRA adapters, hidden triggers | Report, software bill of materials profile, recommendations | 4–6 weeks | Number of components and datasets, depth of review, required standard of evidence |
| AI compliance audit | Assessment against Russian regulation and the EU AI Act: transparency, documentation, human oversight | Compliance map and gap remediation plan | 4–10 weeks | Number of systems, applicable jurisdictions, availability of internal documentation |
| OSINT research and due diligence | Counterparty integrity assessment, information-leak risk analysis, digital footprint review | Findings report with risk assessment and annexes | 2–5 weeks | Depth of review, number of subjects, required standard of evidence |
| Guardrails implementation | Inter-agent authentication, memory isolation, oversight loops, hardware kill switch | Working protection layer and operating procedures | 2–4 months | Number of agents and tools, access mode, production requirements |
| Public outreach | Lectures, webinars and workshops for executives and engineers: risks of agentic systems, defensive practice. No educational programmes are run and no documents of education are issued | Team that attended the series; teaching materials | by agreement | Format (webinar or workshop), number of participants, volume of materials |
| Bug Bounty Automation | Continuous vulnerability-discovery programme: permitted scope, report intake, verification that findings reproduce | A running programme, a disclosure policy, a public report without exploitation details | launch 2–6 months | Programme scope, number of researchers, disclosure requirements |
| AI adoption concepts for an industry | Process breakdown step by step, data map, use cases with an estimate of effect and cost | Concept, roadmap and the boundaries of the work | 4–8 weeks | Industry, depth of the review, volume of data, regulator requirements |
Who we are for, and who we are not for
The Institute measures and tests — it does not implement. The work is narrow, and not everyone needs it. We say so up front: it is fairer for you and for us.
Come to us if you need
- to check someone else’s AI system or contractor — does it work as claimed, and where are the limits of that claim;
- to understand how agents behave in a multi-agent system: alone, together, and when one passes a command to another;
- an internal regulation or standard for AI use — with testable criteria, not slogans;
- a report that survives a question from a regulator, an investor or a court: method, protocols, versions and dates;
- a specific dispute resolved — a detector verdict, a leak, the behaviour of a model.
Do not come to us if you need
- a chatbot built or a CRM implemented turnkey — that is deployment, integrators do it;
- a licence, a certificate or an “approval” — we do not issue permits;
- a vendor who will sell and install a defence product — we neither sell nor install it: that is what independence of assessment is.
If you are not sure, write to us. We will say plainly whether this is our work or not, and who to go to if it is not.
More on selected services
Four areas that generate most enquiries. For each we set out what the work includes, what the client receives and where our limits are.
Guardrails: oversight loops instead of a single filter
Guardrails are not one input filter and not a list of banned words. They are several independent loops, each aimed at its own class of failure: input and output filtering, intent checks before a tool call, mutual authentication between agents, memory isolation and emergency shutdown. We design them for a specific architecture rather than selling a ready-made box.
- Input and output filtering and classification, including indirect injections in documents and on web pages
- Intent checks before tool calls and before requests to external services
- Mutual authentication of models and agents: an instruction from another agent is rejected unless its source is verified
- Memory and log isolation: a shared memory file is the most common infection channel in multi-agent systems
- An oversight loop: an independent observer that can see the decision chain and stop it
- Emergency shutdown — software and hardware, tested against a real scenario rather than on paper
Timeline 2–4 months. Price is scoped individually; factors — number of agents and tools, access mode, production requirements.
A limitation we state before work begins: guardrails reduce likelihood and impact but do not guarantee safety. A system with access to tools remains a system with residual risk; the report names the risk that remains and why.
Compliance audit: Russian regulation and the EU AI Act
We assess a system against requirements already in force and against those about to take effect. The Russian track: Federal Law No. 243-FZ of 26 July 2026, requirements for large foundation models, and notification of rights in generated content. The European track: Regulation (EU) 2024/1689 and its application timetable by risk level. For organisations operating in both markets the two tracks are separated within a single compliance map.
- Inventory of AI systems and classification of each by risk category
- A requirements map: what is mandatory now, what is coming and by when
- Model documentation: purpose, data, limitations, known failures
- Human oversight: where it is mandatory and how its presence is evidenced
- Transparency and marking of generated content
- A gap remediation plan with an effort estimate for every item
Timeline 4–10 weeks. Price is scoped individually; factors — number of systems, jurisdictions, availability of internal documentation.
We do not issue a “fully compliant” conclusion: compliance is confirmed as of a date and within the scope examined, and the report says so plainly. Legal decisions and their consequences remain with the client.
OSINT and digital footprint
Open sources give more than is commonly assumed: team composition, technology stack, contractors, leak history and traces of automation can be collected without touching closed systems. We do this for two purposes — assessing a counterparty before a transaction and assessing a client’s own risk surface.
- Organisation profile: affiliations, officers, participation in procurement, litigation
- Technical footprint: domains and subdomains, certificates, exposed services, infrastructure traces
- Leaks and mentions: breach datasets, forums, repositories, openly accessible storage
- Public digital footprint of employees that is relevant to social engineering
- Risk assessment and remediation priorities — from the cheapest to the most expensive
Timeline 2–5 weeks. Price is scoped individually; factors — depth of review, number of subjects, required standard of evidence.
We work only from lawful open sources and within the scope agreed in writing with the client. Information about individuals is not collected beyond what is necessary.
Bug bounty automation: continuous discovery instead of a one-off review
A one-off review records the state of a system on a given date. A bounty programme turns vulnerability discovery into a continuous process: external researchers work to published rules, findings arrive in a stream, and the Institute handles intake, verification and prioritisation.
- Programme design: permitted scope, reward levels, disclosure rules
- A responsible disclosure policy: timelines and exceptions for critical infrastructure
- A report intake platform and verification that findings are reproducible
- Filtering out false and duplicate reports and rating severity
- Handing findings to engineering and tracking remediation
- A public report at the end of the programme — without exploitation details
Launch 2–6 months. Price on request: it depends on the scope of the programme and the reward pool.
A bounty programme does not replace penetration testing: it draws on a broad community of researchers and finds the unexpected, but it does not give systematic coverage. We say this before work starts, not afterwards.
AI adoption concepts for an industry, with an effect estimate
We set out where AI actually delivers a result in a given industry and estimate the effect before development starts. The client receives not a story about what technology can do, but an examination of its own processes: which tasks are already solvable, which need data that does not exist yet, and which of them pay back.
- Process review stage by stage: where a person decides and where the decision can go to a model
- A data map: what exists, what is missing, what it costs to collect the missing part
- Use scenarios with an estimate of effect and cost of implementation
- Requirements for models and agents: class, mode of operation, deployment perimeter
- Risks: model error, leakage, vendor dependency, regulatory requirements
- A roadmap and a division of work: what the Institute does and what the technology partner takes on
Timeline 4–8 weeks. Price on request: it depends on the industry and the depth of the study.
The Institute produces the concept and the effect estimate; industrial implementation is carried out by a technology partner. That separation keeps the research side independent of contracting interests: we have no stake in proving that implementation is needed — we assess whether it pays back.
Public outreach: formats and content
A series of lectures, webinars and workshops for managers and engineers: risks of agentic systems, breakdowns of real incidents, defence practice. The programme is built around the audience task, not the other way round.
- one to five sessions, matched to your level and your task;
- cases from our chronicle: what happened, who answered and what it cost;
- a workshop: how to set an agent a task safely and how much authority to give it;
- teaching materials that stay with the team;
- format and venue by agreement, on site or remote.
Term and cost depend on format, attendance and the volume of materials. We do not issue educational documents: this is public outreach, not an education programme.
What we do not do
- We do not test systems without the written consent of their owner
- We do not sell or transfer attack tools or attack methodology
- We do not publish vulnerability information before the agreed deadline expires
- We do not take on work whose result cannot be verified by measurement
Software development and IT activity
This section is published in accordance with the requirements applicable to the official website of an organisation operating in the field of information technology in the Russian Federation. The mandatory information in Russian is available on the contacts page.
Fields of IT activity
- Research and experimental development on natural sciences and engineering (OKVED 72.19 — primary code)
- Research and experimental development on social sciences and humanities (72.20)
- Technical testing and analysis: the Institute's testing platform (71.20.9)
- Information security and critical information infrastructure protection activities (74.90.9)
- Software development for analysis and monitoring of multi-agent AI systems (62.01)
- Consulting and computer technology services: AI system audit, guardrails implementation (62.02)
- Testing of AI systems, red teaming, security assessment (62.09)
- Data processing, hosting and monitoring services (63.11)
- Web portals: publishing the Institute's materials and services (63.12)
- Market research and public opinion polling: surveys and studies (73.20)
- Organisation of conferences and exhibitions (82.30)
- Public outreach: lectures, webinars and workshops on AI (no educational programmes are run, no documents of education are issued)
- Publishing of journals and periodicals: the digest and analytical notes (58.14)
- Other publishing activities (58.19)
Codes of IT activity
The list of activity codes in the field of information technology is determined in accordance with order No. 449 of the Ministry of Digital Development of 11 May 2023 and is submitted with an application for state accreditation. The current list of codes is published in the legal information section.
Programming languages and development tools
Python, TypeScript/JavaScript, SQL; frameworks for large language models and multi-agent systems; containerisation (Docker), version control (Git), automated testing, monitoring and logging systems; data analysis and machine learning libraries.
Software rights and licensing
Exclusive rights to software and databases developed by the Institute belong to the Institute or to the rights holder named in the contract. Licensing terms are described on the products page. The Russian-language statement of exclusive rights and licensing is provided in the mandatory Russian section.
Need an assessment of your system?
Send a short description of your architecture — we will propose a scope, timeline and indicative price.
How does an AI system assessment work?
It starts with a short description of the architecture: which models and agents are in play, what tools and data they can reach, and where the trust boundary runs. From that description we name the scope of work and an indicative price, then we run the assessment on our test range and issue a report: what was tested, how, what was found, what risk remains and what to do about it. The work is delivered under a contract with a fixed scope, timeline and result.
What does an assessment cost and what drives the price?
Four things drive the price: the number of models and agents, how many tools they can call, the operating requirements, and whether reporting to a regulator is needed. That is why we publish no price list: two similar-looking systems can need very different amounts of work. Once we have your architecture description, you get scope, timeline and price — before any work starts, with no hidden stages.
How long does the work take?
As a guide: an express assessment of one system takes 2–4 weeks, a compliance audit 4–10 weeks, and implementing oversight loops (guardrails) 2–4 months. The timeline is fixed in the contract. If the scope has to change during the work, we say so before the change, not after it.
What is in the report and in what form do we receive it?
The report contains the scope of the assessment, the methodology, the list of scenarios executed, the failures found with reproduction steps, an assessment of residual risk and remediation advice. We deliver it as a PDF and in a form suitable for handing to your engineering team. Vulnerability details are not published before the agreed period expires — we keep to that after the work ends as well.
Is the organisation registered yet — can we sign a contract?
The Institute is still in the process of state registration: it has no company registration number or tax number yet, so we do not sign contracts and do not take on delivery obligations. This is stated plainly on the «About» and «Access to prices» pages. Research, publications and open data do not depend on registration and are already under way; we will announce separately when we are ready to take orders.