The corporation realized that AI should not yet be trusted with complex tasks.
Meta* experienced an unfortunate incident involving its own artificial intelligence. An internal bot gave an employee incorrect advice, which nearly led to a data leak .
The incident occurred last week and lasted nearly two hours. During this time, company employees gained access to unauthorized user and service information. Meta spokesperson Tracy Clayton stated that user data was ultimately unaffected.
The problem began with a simple question on an internal forum. An engineer decided to investigate the issue using an internal AI agent , which the company compares to OpenClaw and uses in a secure development environment. After analyzing the issue, the bot not only prepared a response but also published it in an open thread without approval. The response was intended for only one employee.
Another employee followed the advice in this post. The recommendation proved inaccurate and resulted in a SEV1 incident, one of the most severe in Meta's rating system. As a result, employees temporarily gained access to sensitive data without proper permission. The issue has since been resolved.
Meta emphasizes that the AI agent itself performed no actions. The bot merely posted a response, albeit an erroneous one. According to Clayton, a human could have made the same mistake, but would likely have verified the information first and reached a more informed conclusion. The employee interacting with the system understood they were dealing with an automated bot. A warning to that effect was included on the page.
A similar incident occurred a month ago, also related to OpenClaw. An employee asked an AI to parse emails, and the bot began deleting messages without permission. Such systems are designed to operate autonomously , but they can still misinterpret commands and return erroneous responses. This is the second time Meta has encountered the consequences of such failures.
Meta* experienced an unfortunate incident involving its own artificial intelligence. An internal bot gave an employee incorrect advice, which nearly led to a data leak .
The incident occurred last week and lasted nearly two hours. During this time, company employees gained access to unauthorized user and service information. Meta spokesperson Tracy Clayton stated that user data was ultimately unaffected.
The problem began with a simple question on an internal forum. An engineer decided to investigate the issue using an internal AI agent , which the company compares to OpenClaw and uses in a secure development environment. After analyzing the issue, the bot not only prepared a response but also published it in an open thread without approval. The response was intended for only one employee.
Another employee followed the advice in this post. The recommendation proved inaccurate and resulted in a SEV1 incident, one of the most severe in Meta's rating system. As a result, employees temporarily gained access to sensitive data without proper permission. The issue has since been resolved.
Meta emphasizes that the AI agent itself performed no actions. The bot merely posted a response, albeit an erroneous one. According to Clayton, a human could have made the same mistake, but would likely have verified the information first and reached a more informed conclusion. The employee interacting with the system understood they were dealing with an automated bot. A warning to that effect was included on the page.
A similar incident occurred a month ago, also related to OpenClaw. An employee asked an AI to parse emails, and the bot began deleting messages without permission. Such systems are designed to operate autonomously , but they can still misinterpret commands and return erroneous responses. This is the second time Meta has encountered the consequences of such failures.
