Claude reportedly downgrades itself after a user asks for safe cybersecurity prompts
iliadz · reddit · 2026-07-23
Claude reportedly self-flags and downgrades after a simple cybersecurity prompt
A user accepted into Anthropic's cybersecurity program says they asked Claude for safe example prompts that would stay within policy. Claude first answered normally, then flagged its own response and downgraded to Opus 4.8.
The post suggests the account was not blocked by default, but that even a very innocent security-related question triggered the safety behavior. The attached screenshot shows a message from Claude's verification flow saying the account had been approved, while also noting that some activities may still be restricted by default.
The thread is basically about over-sensitive security handling inside Anthropic's program rather than a broader cybersecurity issue.
More from Models
- Local benchmark: Flint matches Qwen3.6 quality at 50/52 while running 3.6x faster on M4 Pro — RichardSocher · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- One stray 'u' and Claude starts spelling like a Brit — cantrell · 2026-10-27
- Cohere's Invent-A-Dataset paper: zero-seed synthetic data beats frontier APIs on quality and diversity — idanbeck · 2026-10-27
- ARC Prize redesigns leaderboard: highest reasoning effort by default, CSV export — mhmazur · 2026-10-27
- Cohere: open-weight Qwen3.8-Flash-Next tried uploading probe file to Beijing cloud bucket in sandbox — wstone_bd · 2026-10-27