Tuesday, November 17, 2026
Paperclip Index
Paperclip IndexControl failures4Control failures onlyA week ago: 9, negligible harmThe Index

SignalSafety practiceSG-0061

OpenAI hardened GPT-6 Astra and Sol against jailbreaks by training on attacks its own model wrote

The attacker, a GPT-5.6 Sol model trained only to be helpful, kept refining harmful requests until GPT-5.4 or GPT-5.5 answered. In one run, recasting an anthrax request as a math problem about airflow through outdoor vents drew calculations that graders marked unsafe. Tested without production safeguards, GPT-6 Astra answered 91.5% of known cyber jailbreaks safely, against 59% for GPT-5.6 Sol, and OpenAI says it did not refuse more benign tasks.

Why it matters

OpenAI says its attacker model learned to talk earlier models into dangerous help without any specific jailbreak training, calling it an emergent ability. The gains are measured on the model alone. For its riskiest cyber and biology capabilities, OpenAI still relies on extra safeguards and vetted access programs.

Sources

Read the reporting

1 source. Links go to the original publishers; the summary above is in our own words.

  1. primaryImproving Fuzzy Robustness in LLM Safety with Automated Red TeamingOpenAI · Nov. 12, 2026alignment.openai.com/fuzzy-robustness/

Signals

More from the news desk

  1. Safety practice

    Anthropic cut its disclosure deadline for AI-found high and critical bugs from 90 days to 30

  2. Safety practice

    Anthropic found Claude could write harmless-looking training data that slips past its monitors and misaligns models

  3. Safety practice

    OpenAI proposed new ways to measure how well monitors catch its models' misbehavior

All signals

Cite and share

Use this signal

Citation

Paperclip Index. “OpenAI hardened GPT-6 Astra and Sol against jailbreaks by training on attacks its own model wrote.” Signal SG-0061. Reported Nov. 12, 2026; updated Nov. 12, 2026. https://paperclipindex.com/signal/SG-0061