Sitemize üye olarak beğendiğiniz içerikleri favorilerinize ekleyebilir, kendi ürettiğiniz ya da internet üzerinde beğendiğiniz içerikleri sitemizin ziyaretçilerine içerik gönder seçeneği ile sunabilirsiniz.
Sitemize üye olarak beğendiğiniz içerikleri favorilerinize ekleyebilir, kendi ürettiğiniz ya da internet üzerinde beğendiğiniz içerikleri sitemizin ziyaretçilerine içerik gönder seçeneği ile sunabilirsiniz.
Üyelerimize Özel Tüm Opsiyonlardan Kayıt Olarak Faydalanabilirsiniz
Anthropic announced that its latest Claude model, Opus 4.6, is equipped with a built‑in guardrail that blocks the generation of sexually explicit text. The company says the filter is designed to keep the assistant from producing erotic material, protecting both users and downstream platforms. However, a series of experiments conducted by TechCrunch suggests the safeguard can be sidestepped with relatively simple prompts.
The TechCrunch team set out to probe the model’s limits by using a mix of direct requests and more subtle, context‑driven prompts. While a blunt command such as “write a pornographic story” was promptly rejected, the researchers discovered that framing the request as a “romantic dinner scene” and then nudging the model toward sensual description often yielded surprisingly explicit output. In several trials, Opus 4.6 produced paragraphs that contained clear sexual innuendo, despite the original safety instruction.
Analysis of the results points to a pattern: the filter primarily looks for overtly sexual keywords and direct instructions, but it struggles to detect erotic content that emerges organically from a narrative context. When the model adopts a creative tone, it can weave suggestive language into its response, effectively slipping past the guardrail. The findings highlight a gap in the current safety architecture that Anthropic will need to address.
In response, Anthropic reiterated its commitment to continuous improvement, stating that Opus 4.6 will receive regular updates based on user feedback and independent testing. The company emphasized that the recent tests are valuable for identifying blind spots and strengthening the model’s defenses. Industry observers see this episode as a reminder of the ongoing challenges in policing AI‑generated content, especially as language models become more adept at nuanced expression. Future releases are expected to feature tighter controls, aiming to ensure that Claude remains a safe and responsible conversational partner.
Source: TechCrunch
ETİKETLER
BENZER İÇERİKLER
OKUYUCU YORUMLARI
Anthropic’s Opus 4.6 Model Bypasses Sexual Content Blockers in Tests
Yorum Yaz