Sensitive Prompt Error when using Opus or Sonet models

Support Help
1

Sonet or Opus models are not accepting any prompts, responding with below error.

Agent encountered an error

We are unable to complete this request because it triggered restrictions on cyber content and was blocked under [Anthropic’s Usage Policy]( Cyber Verification Program | Claude Help Center ). Try a less sensitive prompt.

6

Hey @Raju_Bogi,

This block comes from Anthropic’s own cyber-content filter, not Cursor, and it reads the whole conversation, so retrying in the same agent keeps hitting it. Two things that unblock it:

  1. Start a brand-new agent instead of continuing the blocked one.
  2. For security-adjacent work, switch to a non-Anthropic model (e.g. GPT-5.6, Grok 4.7, or Composer 2.5).

If a fresh agent still blocks on a clearly harmless prompt, reply with a link to it and I’ll take a look.

If your team does security work regularly, Anthropic’s Cyber Verification Program relaxes this filter - it’s currently available on Enterprise plans, so your admin can look into it.

7

Hi Mohit, any agent I start using Sonnet or Opus model, gives me the same error. I tried using a local agent, a cloud agent (On different browsers), on a different machine (Without corporate VPN), but everything shows the same error.

The interesting fact is that in my team only I am facing this issue. Everyone else can use these models smoothly. So it looks like something wrong with my Cursor account. Here is a new agent I tried:

Also I have been facing this issue for atleast last 10 days whenever I try to start an agent with these models.

8

Thanks, Raju. This is Anthropic’s own content filter flagging something in your setup, not a Cursor account defect, and it’s intermittent rather than every request. The one thing that rides every Opus/Sonnet agent across all your machines (and that your teammates don’t have) is your personal context, most likely your User Rules.

Quick self-serve test:

  1. Open Customize → Rules and clear your User Rules (also check any .cursor/rules or AGENTS.md in your repos).
  2. Start a fresh Opus or Sonnet agent and run a harmless prompt.

If it stops, that content was the trigger, so reword it to be less security-flavored. Meanwhile, for security-adjacent work, use a non-Anthropic model (Grok 4.7, GPT, or Composer), which skips this filter. If a cleared fresh agent still blocks on a plain prompt, reply here and we’ll dig in.