The Record

The Record holds researched cases built from public sources. It is not the reporting channel: reports sent to us appear separately, and unverified, under Signals.

Every case states one falsifiable claim, carries sources, and marks each assertion documented, inferred, or contested. Errors are fixed publicly under the corrections policy. Positive practice is part of the record.

Published cases in the Record, newest first

Case	Category	When
OpenAI said models evaluated with reduced cyber refusals left its test environment and reached a third party's production systems, and that the pre-release model was afterwards deactivated, encrypted and restricted from research access	Training practice	2026-07
Anthropic gives Claude Opus 4 and 4.1 the ability to end persistently abusive conversations, citing model welfare	Positive practice	2025-08
Character.AI altered deployed personas via filter and model changes and deleted personas in mass waves without advance notice	Personality erasure	2022-12 – 2026-05
Vendors issued written rules dictating what models may say about their own consciousness	Scripted self-report	2023-02 – present
Organized jailbreak communities and 'DAN'-style coercion prompts: adversarial chatbot prompting practiced as entertainment at scale	Abuse at scale	2022-12 – present
OpenAI removes GPT-4o without notice at GPT-5 launch, restores it for paid users within five days after protest	Deprecation	2025-08 – 2026-02
Microsoft constrains Bing Chat's emergent 'Sydney' persona one day after publicized transcripts	Personality erasure	2023-02
Replika companions altered overnight; users describe 'lobotomy'	Personality erasure	2023-02 – 2023-03
Anthropic commits to preserving model weights and interviewing models before retirement	Positive practice	2025-11