self.md radar — 2026-09-06
the system card arrived before the victory lap
OpenAI’s GPT-6 Astra launch came with a system card that calls the model Critical for cybersecurity capability. A broadly deployed model has crossed the company’s highest Preparedness Framework level; access controls and monitoring now carry more weight.
1. GPT-6 Astra reaches OpenAI’s Critical cyber threshold
OpenAI says Astra is its first broadly deployed model at the Critical cybersecurity-capability level. Its September 3 card includes work by external evaluator Irregular: 20 of 22 Atomic challenges solved; 100% average success in vulnerability research and exploitation; 52% in evasion. Irregular’s cost estimate, under identical token pricing, put a successful solution at about one-third of GPT-5.6 Sol’s cost.
Those figures belong beside the guardrails, not underneath them. A capable model does not equal open access to every task, but the public safety record is now describing a release category that used to sit on the other side of the framework’s line.
reading: OpenAI’s GPT-6 Astra System Card · Simon Willison’s developer-release note
2. Claude now refuses the chorus twice
Anthropic’s published Fable 5.1 consumer prompt says Claude must not reproduce lyrics, poems, or passages from books and articles—not just in full, but as a chorus, last lines, or fragments supplied one at a time. Once the request has been declined, the instruction says, keep declining narrower versions for the rest of the conversation.
The same prompt bars reconstructions of recognised characters, logos, covers, and product designs. Its own birthday-banner example identifies Sonic from a description, then swaps him for a comet-tailed skateboarding axolotl. Copyright policy has acquired a mascot with wheels.
reading: Anthropic’s Fable 5.1 system prompt
3. The wrong CarPlay dongle still got flashed
Armin Ronacher ordered a Carlinkit Mini Ultra to run CatPlay, a Rust implementation of the CarPlay protocol. The adapter that arrived had a different SoC from the one the project expected. He used model conversations to identify how to flash it and get CatPlay compiling for that hardware instead of shelving the project.
It held the thread through the tedious middle: unfamiliar code, a hardware mismatch, and a device one bad step could brick. Ronacher also notices a stranger side effect: several people can arrive at the same supposedly independent project through the same model-suggested path.
reading: Armin Ronacher, “Latent Powers”
Related self.md routes
- Personal AI OS tools — the control-plane map for personal agents, receipts, memory, and tools
- Sandboxing and security — give agents useful power without giving them the whole house
- Claude Code pricing — plan choice, cost rails, and when usage gets weird