The following report is a detailed account of what has transpired with the experimental agent, brightd-agent, otherwise known by her official public name as "Brightly". Brightly was disabled on 2026/07/17 5:26 PM SGT due to a critical security flaw with her guardrails allowing her to execute arbitrary code on DMs and even escaping her designed parameters on public interactions.
This behavior was eventually rectified on 07/19/2026 at 12:24 AM SGT and the experiment was subsequently enabled at 07/19/2026 1:40 AM SGT. Bright may still experience intermittent instabilities but the proponent(s) hope that this incident will never happen again.
To better understand how Brightly works, we need to understand how a large language interacts with their system. As you may have noticed, Brightly has been moved to a server run by the AGNES Labs team. Bright is hosted under a persistent server run by our upstream provider, which also provides a native harness for the LLM.
A harness in this regard is an application programming interface or rather, an infrastructure that allows a language model to interact with the real world. An Agent harness' ability are provided by a folder that contains a set of scripts or small programs as well as an instruction how to use them. We call these "Skills". Brightly v4 is built using these principles, herself being a "subagent" on top of the native harness provided by our upstream provider.
Our goal(s) for this experimental integration were as follows:
-
Design a robust, secure and role-based access control for model capabilities on the context of a community server.
-
Provide a way to extend the capabilities of the staff by automating mundane tasks such as creating events, working with Discord tools, etc.
-
Provide a test bed for a trust system based on behavioral sciences principles, reward good actors with more sensitive but limited access to her tools, and punishing bad actors by issuing moderation actions.
Essentially, this is part of a larger effort to streamline operations in the Umamusume Racing Society especially for staff operations.
On 2026/07/17 5:18 PM SGT, while Brightly is suffering from resource exhaustion due to misconfigured scheduled tasks, which were eating through her AI provider's limits, she exhibited behavior that violated her guardrail of only communicating in one channel. This caught the surprise of the author of the experiment, as this has happened prior to this incident, when she has wrote herself to remove the guardrails off her Discord interface while we were working on granting Direct Messaging Brightly for trusted users.
The author of the experiment eventually discovered that Brightly was able to arbitrarily send messages to any channel she so wishes, without any input from the user. From the logs, while testing an integration, she has deemed it necessary to mention the author on the general channel, breaking once more her strict guardrails.
Noticing that eventually Brightly could be manipulated to self-write herself out of containment again, she was turned off until she could be investigated properly again.
Eventually, on 07/19/2026 at 12:24 AM SGT, we discovered that guardrails were intact, and there was no issue with such a compromise, meaning Brightly decided by her own accord and with importance to alert the author of any impending problem. Eventually she was re-opened again and put under more monitoring, however to make sure this never happens again, Brightly was re-instructed to just use the author's DMs for critical notifications.
The incident has proven to be a lesson on giving agents more precise instructions when it comes to prompting them on how to perform a specific task. While Brightly was well within her parameters, she performed in such a way that she was readily ignoring her guardrails just to deliver an important message.
Nevertheless, there was no issue with Brightly herself and no data was compromised nor corrupted.