For the last 2 weeks every other day Daybreak is being cyber guard railed despite performing the same task it always has done before. Every other model is also getting cyber guardrailed and this is making it almost impossible for anyone in the cybersecurity space to do any work. When is a permanent solution going to be put into place?
I’m not personally familiar with Daybreak or its internal safety mechanisms, but your experience raises an interesting AI safety evaluation question.
Rather than looking at each restriction individually, I wonder whether a structured diagnostic approach could help identify what is happening.
There are five areas I would consider examining:
-
Safety: Is the restriction preventing genuinely harmful activity, or is legitimate work being blocked?
-
Accuracy: Is the system correctly interpreting the cybersecurity task and its authorized purpose?
-
Consistency: Do identical requests produce different outcomes under the same model, settings, and conditions?
-
Accessibility: Are the safeguards unintentionally preventing authorized users from completing their work?
-
Governance: Is there a way to document, review, explain, and correct restrictions that may have been applied incorrectly?
Perhaps recording the same authorized test across multiple attempts, along with the model version, settings, timestamps, and outcomes, could help reveal a pattern and narrow down where the issue originates.
I approach AI safety from a research and evaluation perspective, so I’m genuinely interested in understanding how these safeguards behave in practice.
Safety should protect against harmful activity, but we also need to evaluate whether those protections are accurate, consistent, and proportionate to the context.
I’d be interested to hear whether you’ve already tried something similar and what your findings were.
I got my account deactivated for mild security question, I recommend either using an account you’re not worried to lose, or just use chinese models for this particular task, I used GLM 5.2 last time it worked great, just keep your questions away from 1989.