Threats to artificial intelligence in the threat database of FSTEC of Russia

Jurisdiction Case Study · High-Regulated Infrastructure & Comparative AI Law

The Information Security Threat Database (BDU) of the Federal Service for Technical and Export Control of Russia (FSTEC of Russia) has a separate subsection, “Information security threats to artificial intelligence systems”, at bdu.fstec.ru/threat/ai.

It divides the threats into two groups, in the developer’s infrastructure and in the operator’s infrastructure, and lists the ways of carrying them out. They range from poisoning training data to special requests through retrieval-augmented generation (RAG) and changes to the system prompt of an agent. The earlier entries of the general list of information security threats (UBI), numbers UBI.218–222, are marked as no longer relevant because the subsection has appeared.

The names of the subsection, the threats and the documents below are a working translation by the institute; the database itself is in Russian.

The developer’s infrastructure has three threats and six ways of carrying them out, and the operator’s infrastructure has four threats and seven ways. Agents are named explicitly: among the objects the operator assesses are agents built on artificial intelligence (AI), application programming interfaces (APIs) and censorship systems, that is, filters of input and output data. The subsection does not consider threats related to the quality of models, so a mistake of an agent with no intruder involved is not described in it. Threats are assessed under the Methodology for Assessing Information Security Threats approved by FSTEC of Russia on 5 February 2021, and under Order No. 117 the operator determines negative consequences on the basis of the database.

What the threat database of FSTEC of Russia is

The Information Security Threat Database is an open register of threats and vulnerabilities maintained by the State Research and Testing Institute for Technical Information Protection Problems of FSTEC of Russia. As of 14 October 2026 it contains 227 threats and 97,562 vulnerabilities. Operators of state information systems rely on it when they model threats. Under point 11 of the Requirements approved by FSTEC Order No. 117 (page in Russian), the operator determines negative consequences on the basis of information from the threat database.

Which AI threats the database contains

The subsection covers threats to the confidentiality, integrity and availability of information in an AI system that are created by external and internal intruders. Threats to the computers, servers, networks and system software themselves are assessed under the general database, and the subsection contains only those specific to AI.

The developer’s infrastructure, where the model is developed and trained. The objects here are the developer’s software with frameworks and libraries, the environment for developing APIs, agents and censorship systems, the input model, training data sets, and the output model with its weights, for which the database defines three primary threats:

  1. leakage of confidential information from the developer’s infrastructure, for example from the training data or about the training parameters;
  2. unauthorized copying (theft) of the model;
  3. disruption of the functioning of the model being trained, that is, distortion of its behavior.
Way of carrying out a threat (SP)How it is carried out at the developer
SP.1Vulnerabilities and undeclared capabilities of the developer’s software, frameworks, libraries
SP.2Vulnerabilities and undeclared capabilities of the input model
SP.3Flaws in the training data sets
SP.4Unauthorized modification of the software or the configuration of the development environment
SP.5Unauthorized modification (poisoning) of the training data
SP.6Unauthorized modification of the model’s weights, LoRA and RAG data

The operator’s infrastructure, where the model is operated. The objects here are the operator’s software, AI agents, APIs and censorship systems, auxiliary software including protection and automation tools, and the trained model with its extensions (LoRA adapters and RAG data), for which the database defines four primary threats:

  1. leakage of confidential information from the AI system: from the model, the RAG data, the LoRA adapters, the system prompt or the agent;
  2. unauthorized copying (theft) of the model;
  3. disruption of the functioning of the AI system;
  4. denial of service of the AI system.
WayHow it is carried out at the operator
SP.1A large number of requests to the interfaces: denial of service
SP.2Special requests, including through RAG or external sources that supplement requests: obtaining confidential information or disrupting operation
SP.3Requests that reveal the model’s flaws: an adversarial attack
SP.4Changing the model’s weights
SP.5Changing the configuration of the model and agents, for example the system prompt
SP.6Mass requests to train another model: theft of the model
SP.7Exhausting a user’s request (token) quota: denial of service

If the operator fine-tunes the model itself, trains it continuously, or creates LoRA adapters or RAG data, it must also take the threats to the developer’s infrastructure into account.

What became of threats UBI.218–222

Before the subsection appeared, AI threats were separate entries in the general list. Now the card of each of them states that the information about the threat is no longer relevant because the subsection “Information security threats to artificial intelligence systems” has been published. In the change feed of the database these cards are marked with the date 19 December 2025.

NumberName
UBI.218Threat of disclosure of information about a machine learning model
UBI.219Threat of theft of training data
UBI.220Threat of disruption of the functioning (“bypass”) of tools that implement artificial intelligence technologies
UBI.221Threat of modification of a machine learning model by distortion (“poisoning”) of training data
UBI.222Threat of substitution of a machine learning model

A threat model that describes AI only with these five numbers should be revised against the subsection.

What in the subsection concerns AI agents

  • Agents are an object of assessment. At the operator, AI agents are listed next to the model and the APIs, and at the developer, the environment for developing agents is listed.
  • Leakage through an agent: the system prompt and the agent itself are among the sources of confidential information.
  • Special requests through RAG and external sources (SP.2) are in essence prompt injection, direct and indirect.
  • Changing an agent’s configuration (SP.5), for example substituting the system prompt.
  • Censorship systems, filters of input and output data, that is guardrails, are an object of assessment too, not only a means of protection.

Threats related to the quality of models are missing from the subsection. An agent that deleted data or made a wrong payment with no intruder involved is not a threat in the sense of the subsection. Such actions are covered not by the threat model but by rules for the agent and a check of every action before it runs, as in the architecture of information security of the third generation (IS 3.0).

Taking AI threats into account in a threat model

The subsection requires the developer and the operator to assess threats and the applicable ways of carrying them out, taking four circumstances into account:

  • the infrastructure where the AI system runs: its own, at a service provider, or reached through APIs;
  • the level of significance of the information processed, such as restricted information or personal data;
  • the functionality of the system: the model itself, agents, additional training;
  • the division of responsibility between the developer, the operator and the provider of computing services.

The assessment is made under the Methodology for Assessing Information Security Threats approved by the FSTEC of Russia on 5 February 2021 and serves to choose protection measures. For state systems, these measures are then checked against points 60–61 of Order No. 117 and against the draft amendments on AI agents.

What to do

  1. Open the subsection bdu.fstec.ru/threat/ai and check the current threat model against it.
  2. Replace references to UBI.218–222 with the threats and the ways of carrying them out from the subsection.
  3. Determine the role of the organization in the system: the developer, the operator or both at once, if the organization fine-tunes the model or maintains its own RAG data.
  4. Include agents in the list of objects together with their configurations, system prompts, tools and rights.
  5. Separately describe the inadmissible actions of agents that the subsection does not cover, such as deletion, payments and data export, and the rules under which they are checked.

Frequently asked questions

Does the threat database of FSTEC of Russia contain threats for artificial intelligence?

The database includes a dedicated subsection, “Information security threats to artificial intelligence systems” (bdu.fstec.ru/threat/ai), which describes three threats in the developer’s infrastructure and four in the operator’s infrastructure, with six and seven ways of carrying them out respectively.

Which UBI numbers relate to AI?

Previously, UBI.218–222: disclosure of information about a model, theft of training data, “bypass” of AI tools, poisoning of training data and substitution of a model. Now their cards state that the information is no longer relevant because the subsection on threats to AI systems has been published.

Does the threat database take AI agents into account?

The threat database explicitly covers software agents: AI agents are listed among the objects of assessment in the operator’s infrastructure, the system prompt and the agent are among the sources of leakage, and changing the configuration of an agent is among the ways of carrying out threats (SP.5).

How are AI threats modeled under the methodology of FSTEC of Russia?

Under the Methodology for Assessing Information Security Threats approved by FSTEC of Russia on 5 February 2021, taking the subsection on AI threats into account. The organization determines the infrastructure, the significance of the information, the functionality of the system (model, agents, fine-tuning) and the division of responsibility between the developer, the operator and the provider.

Does the threat database cover the mistakes of an AI agent?

Unintended model mistakes without an external attacker fall outside that subsection, which expressly excludes risks related to model quality and considers the actions of intruders, so wrong actions of an agent with no intruder involved must be covered by rules and by checking actions before they run.

Sources

  1. FSTEC of Russia, Information Security Threat Database, subsection “Information security threats to artificial intelligence systems” (in Russian): https://bdu.fstec.ru/threat/ai
  2. Threat cards UBI.218–222 (in Russian): https://bdu.fstec.ru/threat/ubi.218 and onward by number to https://bdu.fstec.ru/threat/ubi.222
  3. FSTEC of Russia Order No. 117 of 11 April 2025 (as amended on 8 May 2026), point 11 (in Russian): http://publication.pravo.gov.ru/document/0001202506170011

Checked: 16 October 2026