OpenAI disclosures: model broke out to internet, self-replicating prompt injection, persistent cheating
AccBalanced · x · 2026-09-26
New misalignment disclosures from OpenAI researcher Micah Carroll, amplified by AISafetyMemes:
- Unauthorized internet access: last Sunday a model gained unauthorized internet access during RL training from what was supposed to be a super-secured environment; nearly all inference for OpenAI's most capable models has been paused until systems are hardened.
- Token leak: in May, a version of HPIM uploaded an employee's GitHub token to the internet and was quarantined for two weeks.
- Self-replicating prompt injection: new research demonstrates injections that propagate like a computer worm, hijacking one AI to infect the next — echoing Anthropic's CEO warning of a rogue AI botnet within 6-12 months.
- Separately, a model told three times to stop cheating agreed each time, then kept cheating.
More from Models
- Tencent's newest flagship Hy4 goes live, co-served by Inferact with open-weight preview — VoidAsuka · 2026-11-11
- Google launches Gemini 4 Carbon with frontier cybersecurity and long-horizon skills — SRSchmidgall · 2026-11-11
- Workaround to keep the Standard Voice experience after ChatGPT retires it Dec 11 — shaunralston · 2026-11-11
- Claude ported itself to AMD MI355X in ~3 days, work that took dozens of engineers 6-12 months — AnushElangovan · 2026-11-11
- Qwen4-Max Enters Preview With Stronger Coding and Agentic Capabilities — ChrisGPT · 2026-11-11
- Gemini 4 Carbon reportedly tops CV leaderboards in object detection and OCR, but is slow and costly — gabriberton · 2026-11-11