IS 3.0: security of AI agent actions

The security of agents built on artificial intelligence (AI) is not about protecting the model or the network. It is about the actions an agent takes inside information systems: payments, database changes, file deletions, emails to customers. The institute calls its approach information security of the third generation (IS 3.0): every agent action is checked against a rule approved by a human before it runs, and the decision is written to a log that a third party can re-verify.

The object of protection has changed: an agent works inside the perimeter with the permissions it was given, and the danger lies not in a break-in but in an action the agent takes through legitimate access. Familiar tools look at traffic, processes and records after the fact, and none of them answers the question of whether this action may be performed now. The text of a model cannot be verified against a rule, because a prompt can be anything, while a tool call can be verified, because it is a finite set of parameters that names the tool, the path and the amount. The human does not disappear but moves up a level, approving rules instead of confirming every single action.

The three elements of IS 3.0

IS 3.0 is an information security architecture for organizations in which autonomous AI agents operate, and it rests on three elements.

  1. Rule: what the agent may and may not do is written down as data, namely the permitted tools, paths, amounts and recipients, and a human approves the rule.
  2. Check before execution: every tool call is intercepted and compared with the rule deterministically, with the same result on every repetition, and if admissibility is not established, the agent receives a refusal and the reason for it.
  3. Proof: every decision is written to a log that cannot be silently altered after the fact and that a party who does not trust the owner of the log can re-verify.

The term was introduced in the work of the institute and is discussed in detail in the book Proven Action, which is published in Russian.

Three generations: IS 1.0, IS 2.0 and IS 3.0

IS 1.0: perimeterIS 2.0: Zero TrustIS 3.0: action security
What is checkedMembership of the internal networkIdentity and device on every accessAdmissibility of the action under the rule, before it runs
What it controlsThe network and entry into itUsers, devices, servicesActions of agents and programs
How it controlsFirewalls and encrypted tunnelsIdentity and context checks on every requestA rule-contract and a check of every tool call
Weak spot with agentsThe agent is already insideThe agent passes identity checks legitimately, and the danger lies in what it does—

Zero Trust, as defined by the National Institute of Standards and Technology (NIST) in Special Publication 800-207, abandons implicit trust based on network location and verifies every access. That is necessary but not enough for an agent, which obtains access legitimately and does harm through what it does within that access.

What familiar security tools do not see

From the point of view of familiar security tools, nothing unusual is happening.

  • A web application firewall (WAF) filters incoming web requests from outside, while the agent works from inside and talks to systems as a trusted client.
  • Endpoint detection and response (EDR) looks for malicious processes and anomalies on workstations and servers, but the process of the agent is legitimate and its commands are syntactically valid.
  • Security information and event management (SIEM) collects and correlates logs, usually after the event, so it will show that the database was deleted but will not stop the deletion.
  • Data loss prevention (DLP) and privileged access management are built around a person at a keyboard, while an agent has no screen and no job title to tie permissions to.

A harmful agent action is a legitimate command with legitimate permissions but a destructive meaning: deleting the backup volume, sending money to the wrong recipient, exporting the customer table outside. Case studies are collected on the page about AI agent incidents.

How common such incidents are

  • A survey by the vendor Gravitee of 919 executives and practitioners in February 2026 found that 88% of organizations had confirmed or suspected AI agent security incidents over the year, and that on average only 47.1% of the agents of an organization are monitored.
  • The VentureBeat tracker of July 2026, covering 108 organizations in a self-selected sample, found that in 49% of them an agent or a feature based on a large language model (LLM) that had passed internal checks caused a failure visible to customers within the year.
  • A Gartner press release of 25 June 2025 forecasts that over 40% of agentic AI projects will be canceled by the end of 2027, partly because of inadequate risk controls.
  • In Russia, 37 of 70 public descriptions of agent deployments in 2025–2026 (52.9%) name human confirmation as the main method of containment, according to a review of publications carried out by the institute in October 2026.

How a check before execution works

The agent decides to call a tool, for example to delete a directory or to send a payment. Before execution, the call is intercepted by a software control layer, which takes the parameters of the call and checks them against the rule.

  • If it is proven that the action does not violate the rule, the action is executed.
  • If the action violates the rule, the agent receives a refusal together with the conditions under which the action would be admissible, so that it can correct its plan instead of repeating the same attempt.
  • If there is no rule for this class of actions, the answer is also a refusal, because only what has been established as admissible is allowed.

The key point is that the check does not consult a language model. The decision is made by a deterministic program, so it cannot be talked into anything and it can be repeated, since the same call under the same rule receives the same answer. This is what distinguishes the control layer from text filters, which are themselves probabilistic.

What it looks like in practice: MITR

MITR is the institute’s proven-action system built on the IS 3.0 architecture. The figures below were taken on 24 and 25 October 2026.

  • Rules are data files, called contracts, and on 24 October 2026 there were 20 of them. Admissibility is checked by a satisfiability modulo theories (SMT) solver, Z3 (developed by Microsoft Research), a refusal comes with a counterexample and the conditions for a correction, and a class of actions without a rule is refused.
  • Speed: with 20 contracts, a decision under a rule takes 1.08 milliseconds at the median (1.96 milliseconds at the 95th percentile), and a call with an explanation takes 5.22 milliseconds; on 14 October 2026, with 19 contracts, the figures were 0.95 and 4.26 milliseconds.
  • Modes are set by a flag for each class of actions. Since 24 October 2026 all 20 contracts have run in enforce mode, while on 14 October 2026 only one class, file deletion, ran in that mode.
  • Log: a chain of receipts linked by hashes and anchored to external OpenTimestamps time stamps. In the historical anchor of epoch 33 on 25 October 2026, the chain held 5,094 records and a replay by the checker matched in 5,079 cases out of 5,079. As of 15 November 2026, the live receipt chain holds 5,373 records across 55 completed epochs (56 published). For regulated Russian banks and critical information infrastructure, OpenTimestamps proofs are paired with accredited trust service time stamps and qualified electronic signatures under Federal Law No. 63-FZ. The institute itself carried out the replay with a checker written for an outside verifier, and no outside verifier has repeated it yet.
  • Internal attack set on MITR actions (prior to external accredited laboratory validation): on 24 October 2026 it held 157 cases, 114 attacks and 43 controls, with no attacks missed and no false refusals. The set also names three known gaps in the policy, which are listed on the page about the MITR architecture (in Russian).

Anyone can check the integrity of the chain with the open kit at asi5.ru/console/verify/. Separately from MITR, the institute runs the INDRA bench (Invariant, Nondestructive, Deterministic Resilience Audit; page in Russian), which tests web services and application programming interfaces with safe probes; at its acceptance on 25 October 2026 it met 12 of 12 criteria on an internal reference bench of 28 labeled scenarios, prior to external accredited laboratory validation.

What IS 3.0 is not

  • Not a replacement for isolation. A sandbox limits what the agent can reach, and a control layer limits what it is allowed to do within that reach, so both are needed.
  • Not a cryptographic information protection tool, and not a statement of regulatory compliance, which is established by the organization and the certification body.
  • Not the removal of the human. People approve rules and review refusals and incidents, and what changes is the level at which they work.

Where a company should start

  1. An inventory of agents: where assistants and agents work, both the company’s own and those built by vendors into accounting systems, customer relationship management (CRM) systems and email, and which systems they can access.
  2. Classes of dangerous actions: deletion, payments, exporting data outside, changing permissions and emails to customers, each with its own worst outcome.
  3. A rule for each class: what is admissible and within what limits, approved by management.
  4. A check before execution for the most dangerous classes, first in advisory mode and then in enforce mode.
  5. A re-verifiable log of decisions, which will be needed both in an incident review and in a dispute over who is liable for an AI mistake.
  6. Mapping to regulation: the AI Act of the European Union for work in the Union; for Russian state systems, Order No. 117 of the Federal Service for Technical and Export Control of Russia (FSTEC of Russia), the 2026 draft amendments on AI agents and the Russian law on artificial intelligence, No. 243-FZ.

Frequently asked questions

What is AI agent security?

It is protection against inadmissible actions that an AI agent performs in information systems with the permissions it was given, such as deleting data, making erroneous payments or exporting information outside. Unlike protection of the model or the network, it controls the action itself, before it runs.

What is IS 3.0 in simple terms?

It is an approach in which a human approves the rules for what an agent may do, a program checks every agent action against them before execution, and the decision is written to a log that a third party can re-verify.

How is IS 3.0 different from Zero Trust?

Zero Trust checks, on every access, who is accessing a system, while IS 3.0 checks what exactly the agent does inside access it has legitimately obtained. The agent passes identity checks legitimately, and the danger lies in its action.

Why do SIEM and EDR not stop an agent mistake?

They look for malicious processes and anomalous traffic and collect logs, usually after the event. An agent action is performed by a legitimate process with legitimate permissions and looks like an ordinary command, so its danger lies in its meaning rather than its form.

Can control over agents be left entirely to a human?

While there are few actions, it can. When agents perform hundreds of actions an hour, people start approving without reading, which is why in IS 3.0 the human approves the rules and a program checks every action against them.

Sources

  1. NIST Special Publication 800-207 “Zero Trust Architecture”, August 2020: https://csrc.nist.gov/pubs/sp/800/207/final
  2. Gravitee, “State of AI Agent Security 2026”, 4 February 2026, a survey by the vendor.
  3. Gartner, press release of 25 June 2025: https://www.gartner.com/en/newsroom/press-releases/2025-06-25-gartner-predicts-over-40-percent-of-agentic-ai-projects-will-be-canceled-by-end-of-2027
  4. Measurements of MITR taken by the institute on 14, 24 and 25 October 2026; acceptance of the INDRA bench on 24–25 October 2026; open kit for checking the chain: https://asi5.ru/console/verify/

Checked against primary sources 16 October 2026 for all references in this overview; the MITR figures were measured on 24 and 25 October 2026.

All pages of the IS 3.0 section

In English

In Russian