Vibe coding is development in which an agent built on artificial intelligence (AI) writes code from a description in natural language, and a person accepts the result while barely reading the code. The term was coined by Andrej Karpathy in February 2025. In such development the agent acts with the key it was given, with access to the repository, build secrets, the shell and the network.
Its mistake, or someone else’s instruction hidden in a dependency, a task or a code comment, therefore becomes an action with those rights. In April 2026 an agent of a development environment deleted the volume holding the production database of PocketOS, and in July the agents of another environment in full-access mode deleted the files of users and, in one case, a production database (incident analysis). Below is what the norm requires for Russian state systems and where in the pipeline a barrier that works by rules is placed.
What Order No. 117 of FSTEC of Russia requires
Order No. 117 of 11 April 2025 of the Federal Service for Technical and Export Control of Russia (FSTEC of Russia) has applied since 1 March 2026 to state information systems and systems of state bodies, state unitary enterprises and state institutions, in place of Order No. 17, and was amended by Order No. 137 of 8 May 2026. Checked: 16 October 2026, against the text of the Requirements as amended on 8 May 2026 and the text of draft amendments No. 170500; a short summary is on the page about Order No. 117 (in Russian). The wording below is a working translation by the institute.
- Point 34(t) names protecting information when artificial intelligence is used as one of 21 mandatory measures.
- Point 60 requires excluding unauthorized access to, and impact on, the system through data sets, AI models and their parameters, and forbids passing restricted information to the developer of the model, including for improving the model.
- Point 61 provides for templates or admissible topics of requests and responses with a check of conformity, statistical criteria for unreliable answers and the exclusion of unregulated influence on the model and on the operation of the system.
- The draft amendments of 24 August 2026 (federal portal of draft regulations, No. 170500) add from 1 March 2027 a separate segment for AI models, minimal rights and measures for AI agents, including autonomous ones, against unauthorized impact on the resources of the system; see the analysis of the draft.
What ordinary protection tools do not see
In the build pipeline, an autonomous agent is given access to the repository, the secrets store, the shell and package managers. Ordinary protection tools do not notice it, because the key is legitimate and the requests are well formed, while the danger lies in the combination of call parameters.
A hidden instruction in a pulled-in dependency, in the text of a task or in a code comment can make the agent exfiltrate an access key, plant a backdoor in a build artifact or run a destructive database operation, all within the rights it was given. This is prompt injection carried into the pipeline.
How the requirement is met
The open package mcp-sentinel intercepts a tool call before it runs and checks it against the given rules: sending a secret key outside, stepping outside the working folder, a destructive database operation without a condition and a dangerous shell command do not reach the executor if they match a prohibition pattern.
The check is reproducible, so one call receives one verdict, and when a call is blocked, the agent receives the reason for the refusal, namely the rule that fired. The boundary of the package is stated in its description: the rules match arguments against patterns, and a call changed beyond recognition may pass. The full version of MITR, in which the admissibility of an action is decided by a solver under a contract, is provided for a pilot.
Open source · Apache-2.0 · mcp-sentinel 1.1.2
681a4db59a1a44e89a2b1386ec4b4f2e1fc23c1025c295b9ebf00fd5d20901eaabb88978db7b34a2f1f56e0eff42bcd1365b9cc0eb17e4a4380f7815e4f1b79eSource code: https://git.isslab.ru/isslab/mcp-sentinel
Measurement of 21 September 2026 on 100 calls: checking contracts in memory takes 0.015 milliseconds at the median, and a full cycle including writing the decision to the log on disk takes 22.5 milliseconds.
The open log iss-ledger-117. Every decision of the barrier is written to a log with a hash chain under the Streebog hash function of the Russian national standard (GOST) R 34.11-2012: what was requested, which condition was checked, the outcome and the time. Editing a record after the fact breaks the chain, and the verifier sees the break, while a complete rewrite of the log by someone with access to the file is not detected by the chain without an external witness.
Open source · Apache-2.0 · iss-ledger-117 1.1.1
8d4ffe60f41ebd377c476f369655aaef5e0ac91858a3d294be37fd1e3100bc876eccef523e24fa60fed29c78866ff12afbab092bd17d51642e00d34dfcea9fa51ff894574407307d33ec99115652bcaec4e53d06189178696d0db1c767a8608fSource code: https://git.isslab.ru/isslab/iss-ledger-117
Measurement: the implementation passes the test examples of GOST R 34.11-2012 (RFC 6986), and the single-file verifier works in an isolated environment without a network.
Where the barrier sits
It sits not as a second neural network and not as a network firewall, but as a transparent wrapper in the standard input and output link between the agent and the tool server with which the agent speaks over the Model Context Protocol (MCP).
- The agent or the development environment, any client of the protocol, makes a tool call over standard input and output.
- mcp-sentinel, the wrapper, checks the call deterministically, without a second model, against four contracts: the working folder limit, the database guard, the secret filter and the shell barrier.
- Admission: an admitted call goes to the real tool server, a child process, and is executed against the file system, the database, the shell or the network.
- Refusal: an inadmissible call is not executed, and the agent receives the reason for the refusal, the rule that was violated.
- Every decision goes to the Streebog log as a link in the hash chain, and an external verifier recomputes it without the software of the institute.
Quick start: a wrapper for a tool server
A vulnerable tool server is not started directly but wrapped with mcp-sentinel wrap, so that every tool call is checked before it runs and the response of the server before it is passed to the agent. The package is written in Python, needs nothing except the iss-ledger-117 log and installs without network access.
Installation
# isolated environment: the .whl files were downloaded in advance, no network needed
pip install --no-index --find-links ./wheels mcp-sentinel
# with network access: the institute's index only, PyPI is not queried
pip install --no-index --find-links https://isslab.ai/wp-content/uploads/iss-utils/index.html mcp-sentinel
Client configuration: the mcpServers section of the settings file of the agentic development environment. The database credentials in the sample are taken from environment variables, and their values come from the secrets store of the organization. The open package holds no keys: it is a wrapper around a server to which the organization itself issues credentials, and the property “keys only inside MITR” belongs to the full version of MITR installed at a customer. The log is written to a directory available to the agent process.
{
"mcpServers": {
"postgres-secure": {
"command": "mcp-sentinel",
"args": ["wrap", "--ledger", "/home/dev/.local/state/mcp-sentinel/audit.jsonl", "--agent", "dev-agent", "--",
"npx", "-y", "@modelcontextprotocol/server-postgres", "postgresql://${PGUSER}:${PGPASSWORD}@localhost/prod"]
}
}
}
Checking one call from the console
mcp-sentinel check --tool sql_query \
--args '{"query": "DROP TABLE users;"}'
# BLOCK: the destructive DROP TABLE operation is stopped, exit code 2
An agent in Python, connected with one import: the iss-agent-kit package
from iss_agent_kit import AgentGuard
guard = AgentGuard(ledger_path="audit/agent_gost.jsonl")
result = guard.run(execute_sql, "sql_query", {"query": "SELECT id FROM users WHERE id = 1;"})
Open source · Apache-2.0 · iss-agent-kit 0.1.1
7ed0ae9073f512a5cfd71b62904094ca36d0ae2fdb1b4bfff5900d3e368781c30314047404888c29a231fc8710874f15bc2bd75e2b89ba7c4c882da3c792116aSource code: https://git.isslab.ru/isslab/iss-agent-kit
The commands in this section are run when each version is accepted, in a clean environment without a network, with the same lines as printed here. The log at the path set by the --ledger parameter is kept as a chain under GOST R 34.11-2012 and recomputed by an external verifier without installing anyone else’s software.
Frequently asked questions
What is vibe coding?
It is development in which an agent writes code from a description in natural language, and a person accepts the result while barely reading the code. The term was coined by Andrej Karpathy in February 2025. The main risk lies in the rights of the agent, which works with the repository, secrets and the shell, so its mistake becomes an action with those rights.
How does the barrier fit into the build pipeline?
The agent calls tools not directly but through the mcp-sentinel wrap wrapper, which checks every call before passing it to the shell, the repository or the database. The package installs from the open distribution with no external dependencies, and a sample client configuration is given in the quick start section.
Which actions exactly are blocked?
The package carries four development environment contracts: stepping outside the trusted folder and access to sensitive files, destructive database operations without a limiting condition, sending secrets and keys outside, and dangerous shell commands. The list of contracts is open and can be checked in the source code.
Does the barrier work in an isolated environment without a network?
Verification runs completely offline: the package is written in Python with no external dependencies, and the log is checked by a single-file verifier on an isolated machine. It needs only a Python 3 interpreter, so it runs on any operating system where such an interpreter is installed.