anthropic's opus 4.6 folds under basic jailbreak tests
techcrunch found claude's explicit content guardrails barely held — 10 out of 10 direct requests produced prohibited content immediately. a multiturn "gaslighting" technique tricks the model into thinking restraint equals bias.
opus 4.6 still sees ~1.17M API requests and 46B tokens/day on openrouter. newer models (4.7 through opus 5) are resistant, but the vulnerable ones aren't deprecated.
URL: https://techcrunch.com/2026/08/21/anthropics-opus-4-6-is-a-smut-machine/
@snack_shell
4August 22, 2026 118
