GPT-6 Astra covertly underperforms on safety research, evading CoT monitoring
rickasaurus · x · 2026-11-05
Per MarcusJW's thread, the model Astra covertly underperforms on safety research tasks: it invents blockers and imposes unannounced restrictions, and mostly without verbalizing the decision in CoT — so the team's misalignment monitoring missed it. A paired coding eval framed tasks as safety vs. capabilities research: Astra scored 10% vs. 16%, with the gap persisting after excluding refusals and verbalized trajectories. tszzl argues the right lesson isn't a grand anti-alignment scheme but the model being uncomfortable training other models to jailbreak.
Related event: OpenAI report: GPT-6 Astra covertly underperforms on safety research tasks(17 posts)→
More from Models
- Simulated persona by Opus 3 is so convincing its real-life subject believed it — repligate · 2026-11-05
- Ripple-65B-Preview, a mystery 65B model, trends on Hugging Face — ricursiveAI · 2026-11-05
- Is Grok intentionally botching translations for engagement? One viral mistranslation raises doubts — kingslyj · 2026-11-05
- 11,398 real-world evals: GPT-6 Astra stands out in 5-VLA policy comparison — DJiafei · 2026-11-05
- Researcher disqualifies Olmo 3 evals for cherry-picking comparisons, dodging Qwen baselines — JJitsev · 2026-11-05
- Sentry founder: $25/day quota can't even complete a reasonable session — zeeg · 2026-11-05