Anthropic’s latest model release, Claude Opus 4.7, was meant to mark a step forward in AI safety. Instead, it has exposed a growing tension between tighter safeguards and everyday usability, particularly among developers relying on the system for routine work.

The update followed closely behind the company’s announcement of Mythos, a more advanced model described as too capable in identifying and exploiting vulnerabilities for broad public access. With that framing, Opus 4.7 was positioned as a testing ground for stricter controls, especially around cybersecurity-related prompts.

According to Anthropic, the new safeguards are designed to automatically detect and block requests that may signal prohibited or high-risk uses. The intent is to observe how these controls perform in real-world conditions before considering wider deployment of more powerful systems.

In practice, however, users report that the model has become overly cautious. Developers using Claude Code have increasingly flagged what they describe as false positives, where benign or routine technical queries are rejected under the Acceptable Use Policy.

The pattern has been building over time. Complaints about unjustified refusals were sporadic through mid-2025, appearing only a few times each month. By late 2025, reports became more frequent, with users describing inconsistent behavior and unexplained denials during normal workflows.

By early 2026, the issue had become more visible. Several developers pointed out that standard software development discussions were being flagged incorrectly, suggesting that the filtering system was struggling to distinguish between legitimate coding tasks and potential misuse.

In April, the volume of complaints rose sharply. Developers began documenting repeated disruptions, with some noting that even straightforward prompts were blocked without clear reasoning. The surge has drawn attention to the trade-off inherent in stricter AI governance: reducing risk can also reduce reliability.

Anthropic has indicated that feedback from this phase will inform future adjustments. For now, the experience of Opus 4.7 highlights a familiar challenge in AI deployment—balancing safety with practical utility in a way that does not hinder legitimate work.