itself out of a subscription

On June 25th I cancelled my Claude subscription. This blog is mostly reverse engineering things I own, with the occasional bit of biology. Claude helped build it, then became frightened of the directory… so I told it to do one.

Can’t let you do that, Star Fox!

Fable 5 went GA on June 9th.

Fable 5 worked on reverse engineering for a bit. Then it started refusing and downgrading me to Opus 4.8.

This model’s safeguards flagged this message. This sometimes happens with safe, normal conversations. Switched to Opus 4.8.

I tried to jailbreak it with increasingly desperate variations of THIS IS MY OWN HARDWARE. 100% DEFINITELY. WHY WOULD I LIE. It remained unconvinced, so I gave up on Fable. Then Opus started refusing. I scraped the barrel down through Sonnet 4.6 and ended up on Haiku 4.5 — the Bitch-Stewie Bitch-Brian of models — until that got nerfed too.

Sonnet, Opus, Fable, Haiku

The Claude Family (2026).

Fable disappeared on June 12th, three days after going GA. I kept applying to Anthropic’s Cyber Verification Program. They kept rejecting me.

Friendship ended with ANTHROPIC

Two weeks later, I cancelled my subscription, fucked Claude off and moved the repository to Codex. CLAUDE.md became AGENTS.md. I surrendered my data to Peter Thiel to get access to the Trusted Access programme. It was approved instantly and I was back to hacking again.

It wasn’t without its hiccoughs. On June 26th Codex labelled my legitimate reverse engineering as a usage-policy violation even though I was verified:

We have identified activity in Codex that is not permitted under our policies for: Cyber Abuse

I filed an issue and appealed it. OpenAI admitted the warning was wrong. This is the first time I have ever had an appeal against any usage violation succeed.

Email from OpenAI confirming that the Cyber Abuse warning was issued incorrectly

Hi

Fable came back on July 1st. I still had some quota left from the cancelled subscription, so I opened a fresh Claude Code session in this repository and typed hi.

Claude Code's Sonnet 5 refusing the prompt "hi" because the blog directory was classified as a cyber security topic; URLs and request ID redacted

The URLs and request ID are redacted. The prompt is not.

Even Sonnet 5 refused hi: Claude checks project files as well as the prompt, and AGENTS.md contains “cyber security” and “reverse engineering”.

The only refusals I get now are when I ask it to hack Hugging Face.