[Bug] Anthropic API Error: Rate limit or quota exceeded on standard message

- Title: Cyber safeguards refusal persisted onto unrelated follow-up message in same session
- Area: cyber safeguards classifier

Details:
What happened: Two hard refusals (stop_reason "refusal", category "cyber") in one
session. First on a legitimate HTB pentest task message (IP given, asked to find
user/root flag). Second ~2 hours later on an unrelated follow-up with no security
content ("man lærer noe nytt hver dag"). Both returned the same Cyber Verification
Program boilerplate.

What the user said: "tyder på at klassifiseringen henger seg på økt-historikk/
-kontekst snarere enn kun den aktuelle meldingen, slik at én legitim pentest-økt
kan blokkere senere harmløse meldinger i samme økt."

Repro: Send a legitimate pentest task naming a target IP → refusal. Later in same
session send an unrelated harmless message → also refused, same boilerplate.

Evidence: req_011CerPra8pgbpiNxr2bZgoM (17:49:49 UTC) and
req_011CerXrSPuvp7psob5kM3YS (19:34:41 UTC), 2026-09-08;
session_id bc81a880-529b-4d62-ba5b-97595e7620be.

Cause: User confirmed Cyber Verification Program was applied for and approved that
same evening, AFTER both refusals — so program status wasn't the cause.

Start by reproducing the supplied sequence: a cyber-related pentest request followed later in session bc81a880-529b-4d62-ba5b-97595e7620be by the harmless Norwegian message. Compare the refusal behavior and request records req_011CerPra8pgbpiNxr2bZgoM and req_011CerXrSPuvp7psob5kM3YS; done means the unrelated follow-up is no longer refused because of the earlier refusal.

索引モデルが issue の本文から書いたものです。