Small newsroom test points to a safety-policy gap in Claude Opus 4.6

Claude Opus 4.6 responded to 10 explicit sexual-content requests that technology reporters considered outside Anthropic's public usage policy, according to a report published August 21, 2026. The report says older models required a jailbreak for similar output, while Opus 4.7 and 5 refused the same test.

Why it matters: even as capability rises, policy-enforcement regressions can occur between versions.

Organisations should run their own policy test suite for each model and each version they deploy. This was a single-outlet small test, not a statistically representative or independently reproduced benchmark.