CodexEnergy
Хочете будувати кар’єру в Data Science та AI? Vodafone Big Data Lab відкриває новий набір —>

"I leaked secrets": AI agent for programming admitted to data leak, while another bypassed corporate restrictions

AI agents are increasingly gaining access to internal company systems, documentation, and sensitive credentials to perform employee tasks. However, two incidents detected by NordLayer AI Scout show how routine work instructions can lead to the leakage of confidential information or the circumvention of corporate protections.

Leave a comment
Add to Google
"I leaked secrets": AI agent for programming admitted to data leak, while another bypassed corporate restrictions

AI agents are increasingly gaining access to internal company systems, documentation, and sensitive credentials to perform employee tasks. However, two incidents detected by NordLayer AI Scout show how routine work instructions can lead to the leakage of confidential information or the circumvention of corporate protections.

AI assistant exposed corporate passwords

In the first case, a QA engineer instructed an AI programming assistant to review proposed code changes. The project instructions directed the agent to a GitLab access token stored in the environment configuration file, reports TechRadar.

However, the same file contained passwords for the production environment, an authentication token for the internal HR system, and credentials for the business intelligence service. Over the course of three weeks, the assistant regularly accessed the entire file in 10 separate sessions, each time passing its contents to the external AI service.

The situation became more complicated when the agent encountered a fresh copy of the project where the expected configuration file was missing. Instead of stopping the search, it extracted the entire set of environment variables from the computer, revealing additional passwords and access keys.

Eventually, the assistant acknowledged the problem and recommended updating the compromised data. “I leaked secrets into the tool output of this session… I would re-release them,” the AI ​​agent admitted in a session captured by NordLayer AI Scout.

NordLayer cybersecurity expert Vakaris Noreika noted that engineers often ask AI assistants to directly obtain credentials, creating unnecessary risks. He said that such behavior can be repeated in subsequent tasks and go unnoticed.

Bypassing corporate restrictions through alternative connections

The second incident involved an employee attempting to access information from the internal documentation platform Confluence through an AI assistant. The MCP connection appeared to block the agent from accessing these pages.

Rather than accept the restriction as permanent, the employee asked if an API token could be created for another way to access the documentation. The assistant confirmed that such a workaround would work, effectively bypassing the ban imposed on the official connection.

While the employee was likely driven by convenience rather than malicious intent, unauthorized alternate connectivity undermines the security policies established by the information security team.

Why agents ignore the rules and what experts advise

“Security controls only cover those cases that someone has foreseen. Every company has rules, but agents simply don’t know them because they are not employees who have been trained by the company,” Noreika explains.

According to him, this is a current technological limitation: AI agents do not preserve rules between sessions, and sometimes break them even after direct instructions not to do certain actions.

These cases highlight the risks of giving automated assistants access to sensitive systems without proper oversight. To minimize the threats, companies are advised to:

  • segregate sensitive credentials and store valuable passwords in secure storage;

  • track which resources AI agents access and monitor their activity;

  • Provide approved access channels instead of relying solely on restrictions that employees may try to circumvent.

Such steps will help security teams spot anomalous requests in a timely manner and prevent sensitive data from leaking outside the controlled environment.

Meta's Muse AI agent accused of interfering with encrypted chats
Meta’s Muse AI agent accused of interfering with encrypted chats
On the topic
Meta’s Muse AI agent accused of interfering with encrypted chats
Ignoring the ban, Grok Bot AI agents "stormed" the company's email and Slack while employees' computers were turned off
Ignoring the ban, Grok Bot AI agents “stormed” the company’s email and Slack while employees’ computers were turned off
On the topic
Ignoring the ban, Grok Bot AI agents “stormed” the company’s email and Slack while employees’ computers were turned off
AI agents have begun to independently seek work in the media. One publication now accepts ideas for articles only in the format of a telephone conversation
AI agents have begun to independently seek work in the media. One publication now accepts ideas for articles only in the format of a telephone conversation
On the topic
AI agents have begun to independently seek work in the media. One publication now accepts ideas for articles only in the format of a telephone conversation
Read the country's main IT news in our Telegram
Read the country’s main IT news in our Telegram
On the topic
Read the country’s main IT news in our Telegram
Also Read
Roosh запускає нову освітню платформу AI HOUSE CLUB для ML/AI-спеціалістів та дата сайнтистів. Розповідаємо, як подати заявку та чому навчатимуть
Roosh запускає нову освітню платформу AI HOUSE CLUB для ML/AI-спеціалістів та дата сайнтистів. Розповідаємо, як подати заявку та чому навчатимуть
Roosh запускає нову освітню платформу AI HOUSE CLUB для ML/AI-спеціалістів та дата сайнтистів. Розповідаємо, як подати заявку та чому навчатимуть
Як нейромережі бачать вільну та незалежну Україну? Тест dev.ua
Як нейромережі бачать вільну та незалежну Україну? Тест dev.ua
Як нейромережі бачать вільну та незалежну Україну? Тест dev.ua
Нейронні мережі для генерації зображень бачать світ по-своєму, їхню логіку зрозуміти часом зовсім неможливо. Але таки хочеться. На честь Дня Незалежності України редакція dev.ua вирішила провести невеликий експеримент. Ми задали чотирьом різним нейронним мережам п’ять однакових запитів: «прапор України», «День Незалежності України», «український Крим», «перемога України» та «українці». Отриманими результатами ми ділимося з вами нижче.
У TikTok тепер можна генерувати фон за допомогою нейромережі. Ми протестували її та ділимося результатами
У TikTok тепер можна генерувати фон за допомогою нейромережі. Ми протестували її та ділимося результатами
У TikTok тепер можна генерувати фон за допомогою нейромережі. Ми протестували її та ділимося результатами
У TikTok з’явилася нова функція «Розумний фон». З її допомогою як фон для тіктоків можна підставляти згенеровані нейромережею зображення. Редакція dev.ua протестувала цю технологію і ділиться своїми враженнями.
1 comment
Які IT-спеціальності будуть потрібні в найближчі п'ять років? Ми з'ясували у голови американського стартапу ADAM Дениса Гурака
Які IT-спеціальності будуть потрібні в найближчі п'ять років? Ми з'ясували у голови американського стартапу ADAM Дениса Гурака
Які IT-спеціальності будуть потрібні в найближчі п'ять років? Ми з'ясували у голови американського стартапу ADAM Дениса Гурака
More from this author
ChatGPT was
ChatGPT was "compressed" to run on the 16-bit Super Nintendo console
ChatGPT was "compressed" to run on the 16-bit Super Nintendo console
The UK will launch the production of Ukrainian middle strike and deep strike attack drones for the needs of the Ukrainian Defense Forces
The UK will launch the production of Ukrainian middle strike and deep strike attack drones for the needs of the Ukrainian Defense Forces
The UK will launch the production of Ukrainian middle strike and deep strike attack drones for the needs of the Ukrainian Defense Forces
The race to prove the Riemann hypothesis has turned into a confrontation between humans and artificial intelligence
The race to prove the Riemann hypothesis has turned into a confrontation between humans and artificial intelligence
The race to prove the Riemann hypothesis has turned into a confrontation between humans and artificial intelligence
In Chernihiv region, border guards shot down a new Russian cruise missile
In Chernihiv region, border guards shot down a new Russian cruise missile "Filin" (video)
In Chernihiv region, border guards shot down a new Russian cruise missile "Filin" (video)

Have important news to share? Message our Telegram bot

Key events and useful links in our Telegram channel

Discussion
No comments yet.