A research engineer prepping slides for a talk on automated AI research hit an immediate API error when asking Claude to fetch a prior-session Artifact link for simple plot color swaps, with the model citing a usage policy breach despite the request involving nothing more than theme adjustments like blue for Claude and orange for Codex.
@eliebakouch Don’t use an unsafe color
i'm preparing my slides for tomorrow and asked claude to change the color of the (auto research) plots and got blocked by the safety filter, everything is fine :)
The block stemmed purely from the cross-session reference rather than any risky content, underscoring how safety layers can interrupt routine tasks such as UI tweaks during research prep with no official Anthropic details on the triggering logic.
Similar over-refusals have surfaced in other non-obvious prompts across web and code interfaces, leaving researchers without clarity on whether API users see different handling or if adjustments followed the July 1 reports.
Users criticized Claude's safety filter for blocking a simple color change request in AI research plots, such as switching from blue for Claude to orange for Codex.
No Digg Deeper questions have been answered for this story yet.
@xeophon you know the worst part, the color was like blue for claude and orange for codex and just asked in another convo to switch it and send the artifact link 😭
@eliebakouch Don’t use an unsafe color