An employee at Meta granted access to a significant volume of confidential data after following a recommendation from an artificial intelligence system. The data became accessible to other Meta engineers who did not have the appropriate level of clearance.
HOW THE INCIDENT OCCURRED
According to The Guardian, an employee posted a question on an internal forum for engineers, asking for help solving a technical problem.
In response, the system with an AI assistant suggested a solution. Following this, other Meta engineers who did not have the appropriate level of clearance gained access to internal and user data.
HOW LONG THE DATA ACCESS LASTED
Access to this data remained open for around two hours. After the incident was discovered, access was revoked and the problem itself was fixed.
WHY THIS HAPPENED
Experts note that such errors are linked to the specific way AI agents operate. Unlike a human, these systems lack contextual understanding and do not always consider the consequences of their recommendations within complex infrastructures.
Furthermore, the AI relies on a limited set of current instructions, which can lead to erroneous actions when information is insufficient.
CONTEXT: THIS IS NOT AN ISOLATED INCIDENT
Similar incidents have been recorded at other technology companies. In particular, there have been previous reports of glitches when using internal AI tools at Amazon, where employees pointed out errors in code and a decline in work efficiency.
The development of so-called agentic AI — systems capable of performing tasks autonomously — has accelerated in recent months, and companies are actively integrating them into workflows.
THE COMPANY'S RESPONSE
Meta stated that no evidence of improper use of user data has been found. The company emphasised that a human could have given the same erroneous recommendation.
An internal investigation is currently underway, the results of which will determine the circumstances of what happened.
WHAT REMAINS UNCLEAR
The reports do not specify exactly what data was affected or how many employees gained access to it.