Something strange is happening at the intersection of cybersecurity and aesthetics. AI guardrails are blocking legitimate security researchers from using tools like ChatGPT and Claude to probe vulnerabilities, while those same models happily assist bad actors who know how to prompt around restrictions. The guardrail, it turns out, is mostly decorative. A velvet rope that only stops people who respect velvet ropes.
When Safety Logic Becomes Aesthetic Judgment
The parallel to algorithmic culture is precise and under-discussed. Kyle Chayka's argument in Filterworld is that recommendation algorithms don't curate so much as they flatten, collapsing the diversity of human taste into whatever is statistically safe. AI safety filters do the same thing to knowledge: they optimize for the appearance of harmlessness rather than actual harm reduction. A 2025 paper in arXiv CS.CY by Marin-Llobet, Lehr, and Banaji found that language models embody and amplify human cognitive distortions, meaning the bias isn't incidental, it's structural. The model isn't neutral. It has taste, and that taste was trained on anxious, liability-conscious humans. The result is what Kyle Chayka describes as algorithmic homogenization: a system that mistakes the smoothing of edges for the elimination of danger.
The Researcher as the New Pirate
Offensive security researchers now occupy the same cultural position as sample-based musicians after the copyright wars: technically doing something legitimate, structurally treated as criminals. The irony is total. The people most qualified to understand AI's attack surface are being locked out by the very systems they need to study. Meanwhile, the Open Veins of Algorithmic Auditing paper from Galdon Clavell and Magaard documents how AI assessment infrastructure lags worst in the Global South, where deployment is fastest. Guardrails, like taste, are unevenly distributed. The velvet rope has a geography.